Longitudinal and Incomplete Data - KU Leuven · Longitudinal and Incomplete Data Geert Verbeke...

Longitudinal and Incomplete Data

Geert Verbeke

geert.verbeke@med.kuleuven.be

Geert Molenberghs

geert.molenberghs@uhasselt.be

geert.molenberghs@med.kuleuven.be

Interuniversity Institute for Biostatistics and statistical Bioinformatics (I-BioStat)

Katholieke Universiteit Leuven & Universiteit Hasselt, Belgium

http://www.ibiostat.be

Interuniversity Institute for Biostatistics and statistical Bioinformatics Royal Statistical Society, London, 26–27 March 2014

Contents

0 Related References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

I Continuous Longitudinal Data 8

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2 A Model for Longitudinal Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

3 Estimation and Inference in the Marginal Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

4 Inference for the Random Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

Royal Statistical Society, London, 26–27 March 2014 i

II Marginal Models for Non-Gaussian Longitudinal Data 62

5 The Toenail Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

6 The Analgesic Trial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

7 Generalized Linear Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

8 Parametric Modeling Families . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

9 Generalized Estimating Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

10 A Family of GEE Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

III Generalized Linear Mixed Models for Non-Gaussian Longitudinal Data 117

11 Generalized Linear Mixed Models (GLMM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

12 Fitting GLMM’s in SAS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

IV Marginal Versus Random-effects Models 140

13 Marginal Versus Random-effects Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

Royal Statistical Society, London, 26–27 March 2014 ii

V Non-linear Models 152

14 Non-Linear Mixed Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153

VI Incomplete Data 185

15 Setting The Scene . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186

16 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238

VII Topics in Methods and Sensitivity Analysis for Incomplete Data 239

17 An MNAR Selection Model and Local Influence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240

18 Interval of Ignorance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249

19 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268

Royal Statistical Society, London, 26–27 March 2014 iii

Chapter 0

Related References

• Aerts, M., Geys, H., Molenberghs, G., and Ryan, L.M. (2002). Topics inModelling of Clustered Data. London: Chapman & Hall.

• Brown, H. and Prescott, R. (1999). Applied Mixed Models in Medicine. NewYork: John Wiley & Sons.

• Carpenter, J.R. and Kenward, M.G. (2013). Multiple Imputation and itsApplication. New York: John Wiley & Sons.

• Crowder, M.J. and Hand, D.J. (1990). Analysis of Repeated Measures. London:Chapman & Hall.

Royal Statistical Society, London, 26–27 March 2014 1

• Davidian, M. and Giltinan, D.M. (1995). Nonlinear Models For RepeatedMeasurement Data. London: Chapman & Hall.

• Davis, C.S. (2002). Statistical Methods for the Analysis of RepeatedMeasurements. New York: Springer.

• Demidenko, E. (2004). Mixed Models: Theory and Applications. New York: JohnWiley & Sons.

• Diggle, P.J., Heagerty, P.J., Liang, K.Y. and Zeger, S.L. (2002). Analysis ofLongitudinal Data. (2nd edition). Oxford: Oxford University Press.

• Fahrmeir, L. and Tutz, G. (2002). Multivariate Statistical Modelling Based onGeneralized Linear Models (2nd ed). New York: Springer.

• Fitzmaurice, G.M., Davidian, M., Verbeke, G., and Molenberghs, G.(2009).Longitudinal Data Analysis. Handbook. Hoboken, NJ: John Wiley & Sons.

• Fitzmaurice, G.M., Laird, N.M., and Ware, J.H. (2004). Applied LongitudinalAnalysis. New York: John Wiley & Sons.

• Ga lecki, A. and Burzykowski, T. (2013). Linear Mixed-Effects Models Using R.New York: Springer.

• Goldstein, H. (1979). The Design and Analysis of Longitudinal Studies. London:Academic Press.

• Goldstein, H. (1995). Multilevel Statistical Models. London: Edward Arnold.

• Hand, D.J. and Crowder, M.J. (1995). Practical Longitudinal Data Analysis.London: Chapman & Hall.

• Hedeker, D. and Gibbons, R.D. (2006). Longitudinal Data Analysis. New York:John Wiley & Sons.

• Jones, B. and Kenward, M.G. (1989). Design and Analysis of Crossover Trials.

London: Chapman & Hall.

• Kshirsagar, A.M. and Smith, W.B. (1995). Growth Curves. New York: MarcelDekker.

• Leyland, A.H. and Goldstein, H. (2001) Multilevel Modelling of Health Statistics.Chichester: John Wiley & Sons.

• Lindsey, J.K. (1993). Models for Repeated Measurements. Oxford: OxfordUniversity Press.

• Littell, R.C., Milliken, G.A., Stroup, W.W., Wolfinger, R.D., and Schabenberger,O. (2005). SAS for Mixed Models (2nd ed.). Cary: SAS Press.

• Longford, N.T. (1993). Random Coefficient Models. Oxford: Oxford UniversityPress.

• Molenberghs, G., Fitzmaurice, G., Kenward, M.G., Tsiatis, A.A., and Verbeke, G.

(2014). Handbook of Missing Data. Boca Raton: Chapman & Hall/CRC.

• Mallinckrodt, C.G. (2013). Preventing and Treating Missing Data in LongitudinalClinical Trials. Cambridge University Press.

• McCullagh, P. and Nelder, J.A. (1989). Generalized Linear Models (secondedition). London: Chapman & Hall.

• Molenberghs, G. and Kenward, M.G. (2007). Missing Data in Clinical Studies.Chichester: John Wiley & Sons.

• Molenberghs, G. and Verbeke, G. (2005). Models for Discrete Longitudinal Data.New York: Springer.

• Pinheiro, J.C. and Bates D.M. (2000). Mixed effects models in S and S-Plus.New York: Springer.

• Rizopoulos, D. (2012). Joint Models for Longitudinal and Time-to-Event Data.

With Applications in R. Boca Raton: Chapman & Hall/CRC.

• Searle, S.R., Casella, G., and McCulloch, C.E. (1992). Variance Components.New-York: Wiley.

• Senn, S.J. (1993). Cross-over Trials in Clinical Research. Chichester: Wiley.

• Tan, M.T., Tian, G.-L., and Ng, K.W. (2010). Bayesian Missing Data Problems.Boca Raton: Chapman & Hall/CRC.

• Verbeke, G. and Molenberghs, G. (1997). Linear Mixed Models In Practice: ASAS Oriented Approach, Lecture Notes in Statistics 126. New-York: Springer.

• Verbeke, G. and Molenberghs, G. (2000). Linear Mixed Models for LongitudinalData. Springer Series in Statistics. New-York: Springer.

• Vonesh, E.F. and Chinchilli, V.M. (1997). Linear and Non-linear Models for theAnalysis of Repeated Measurements. Basel: Marcel Dekker.

• Weiss, R.E. (2005). Modeling Longitudinal Data. New York: Springer.

• West, B.T., Welch, K.B., and Ga lecki, A.T. (2007). Linear Mixed Models: APractical Guide Using Statistical Software. Boca Raton: Chapman & Hall/CRC.

• Wu, H. and Zhang, J.-T. (2006). Nonparametric Regression Methods forLongitudinal Data Analysis. New York: John Wiley & Sons.

• Wu, L. (2010). Mixed Effects Models for Complex Data. Boca Raton: Chapman& Hall/CRC.

Part I

Continuous Longitudinal Data

Chapter 1

Introduction

. Repeated Measures / Longitudinal data

. Examples

1.1 Repeated Measures / Longitudinal Data

Repeated measures are obtained when a responseis measured repeatedly on a set of units

• Units:

. Subjects, patients, participants, . . .

. Animals, plants, . . .

. Clusters: families, towns, branches of a company,. . .

. . . .

• Special case: Longitudinal data

1.2 Rat Data

• Research question (Dentistry, K.U.Leuven):

How does craniofacial growth depend ontestosteron production ?

• Randomized experiment in which 50 male Wistar rats are randomized to:

. Control (15 rats)

. Low dose of Decapeptyl (18 rats)

. High dose of Decapeptyl (17 rats)

• Treatment starts at the age of 45 days; measurements taken every 10 days, fromday 50 on.

• The responses are distances (pixels) between well defined points on x-ray picturesof the skull of each rat:

• Measurements with respect to the roof, base and height of the skull. Here, weconsider only one response, reflecting the height of the skull.

• Individual profiles:

• Complication: Dropout due to anaesthesia (56%):

# Observations

Age (days) Control Low High Total

50 15 18 17 50

60 13 17 16 46

70 13 15 15 43

80 10 15 13 38

90 7 12 10 29

100 4 10 10 24

110 4 8 10 22

• Remarks:

. Much variability between rats, much less variability within rats

. Fixed number of measurements scheduled per subject, but not allmeasurements available due to dropout, for known reason.

. Measurements taken at fixed time points

1.3 Prostate Data

• References:

. Carter et al (1992, Cancer Research).

. Carter et al (1992, Journal of the American Medical Association).

. Morrell et al (1995, Journal of the American Statistical Association).

. Pearson et al (1994, Statistics in Medicine).

• Prostate disease is one of the most common and most costly medical problems inthe United States

• Important to look for markers which can detect the disease at an early stage

• Prostate-Specific Antigen is an enzyme produced by both normal and cancerousprostate cells

• PSA level is related to the volume of prostate tissue.

• Problem: Patients with Benign Prostatic Hyperplasia also have an increased PSAlevel

• Overlap in PSA distribution for cancer and BPH cases seriously complicates thedetection of prostate cancer.

• Research question (hypothesis based on clinical practice):

Can longitudinal PSA profiles be used todetect prostate cancer in an early stage ?

• A retrospective case-control study based on frozen serum samples:

. 16 control patients

. 20 BPH cases

. 14 local cancer cases

. 4 metastatic cancer cases

• Complication: No perfect match for age at diagnosis and years of follow-uppossible

• Hence, analyses will have to correct for these age differences between thediagnostic groups.

• Remarks:

. Much variability between subjects

. Little variability within subjects

. Highly unbalanced data

Chapter 2

A Model for Longitudinal Data

. Introduction

. The 2-stage model formulation

. Example: Rat data

. The general linear mixed-effects model

. Hierarchical versus marginal model

. Example: Rat data

. Example: Bivariate observations

2.1 Introduction

• In practice: often unbalanced data:

. unequal number of measurements per subject

. measurements not taken at fixed time points

• Therefore, multivariate regression techniques are often not applicable

• Often, subject-specific longitudinal profiles can be well approximated by linearregression functions

• This leads to a 2-stage model formulation:

. Stage 1: Linear regression model for each subject separately

. Stage 2: Explain variability in the subject-specific regression coefficients usingknown covariates

2.2 A 2-stage Model Formulation

2.2.1 Stage 1

• Response Yij for ith subject, measured at time tij, i = 1, . . . ,N , j = 1, . . . , ni

• Response vector Yi for ith subject: Yi = (Yi1, Yi2, . . . , Yini)′

• Stage 1 model:

Yi = Ziβi + εi

• Zi is a (ni × q) matrix of known covariates

• βi is a q-dimensional vector of subject-specific regression coefficients

• εi ∼ N (0,Σi), often Σi = σ2Ini

• Note that the above model describes the observed variability within subjects

2.2.2 Stage 2

• Between-subject variability can now be studied from relating the βi to knowncovariates

• Stage 2 model:

βi = Kiβ + bi

• Ki is a (q × p) matrix of known covariates

• β is a p-dimensional vector of unknown regression parameters

• bi ∼ N (0, D)

2.3 Example: The Rat Data

• Transformation of the time scale to linearize the profiles:

Ageij −→ tij = ln[1 + (Ageij − 45)/10)]

• Note that t = 0 corresponds to the start of the treatment (moment ofrandomization)

• Stage 1 model: Yij = β1i + β2itij + εij, j = 1, . . . , ni

• Matrix notation:

Yi = Ziβi + εi with Zi =

... ...

1 tini

• In the second stage, the subject-specific intercepts and time effects are related tothe treatment of the rats

• Stage 2 model:

β1i = β0 + b1i,

β2i = β1Li + β2Hi + β3Ci + b2i,

• Li, Hi, and Ci are indicator variables:

1 if low dose

0 otherwiseHi =

1 if high dose

0 otherwiseCi =

1 if control

0 otherwise

• Parameter interpretation:

. β0: average response at the start of the treatment (independent of treatment)

. β1, β2, and β3: average time effect for each treatment group

2.4 The General Linear Mixed-effects Model

• A 2-stage approach can be performed explicitly in the analysis

• However, this is just another example of the use of summary statistics:

. Yi is summarized by βi

. summary statistics βi analysed in second stage

• The associated drawbacks can be avoided by combining the two stages into onemodel:

Yi = Ziβi + εi

βi = Kiβ + bi=⇒ Yi = ZiKi︸︷︷︸

β + Zibi + εi = Xiβ + Zibi + εi

• General linear mixed-effects model:

Yi = Xiβ + Zibi + εi

bi ∼ N (0, D), εi ∼ N (0,Σi),

b1, . . . , bN, ε1, . . . , εN independent

• Terminology:

. Fixed effects: β

. Random effects: bi

. Variance components: elements in D and Σi

2.5 Hierarchical versus Marginal Model

• The general linear mixed model is given by:

bi ∼ N (0, D), εi ∼ N (0,Σi),

b1, . . . , bN, ε1, . . . , εN independent

• It can be rewritten as:

Yi|bi ∼ N (Xiβ + Zibi,Σi), bi ∼ N (0,D)

• It is therefore also called a hierarchical model:

. A model for Yi given bi

. A model for bi

• Marginally, we have that Yi is distributed as:

Yi ∼ N (Xiβ, ZiDZ′i + Σi)

• Hence, very specific assumptions are made about the dependence of mean andcovariance onn the covariates Xi and Zi:

. Implied mean : Xiβ

. Implied covariance : Vi = ZiDZ′i + Σi

• Note that the hierarchical model implies the marginal one, NOT vice versa

2.6 Example: The Rat Data

• Stage 1 model: Yij = β1i + β2itij + εij, j = 1, . . . , ni

• Stage 2 model:

β1i = β0 + b1i,

β2i = β1Li + β2Hi + β3Ci + b2i,

• Combined: Yij = (β0 + b1i) + (β1Li + β2Hi + β3Ci + b2i)tij + εij

β0 + b1i + (β1 + b2i)tij + εij, if low dose

β0 + b1i + (β2 + b2i)tij + εij, if high dose

β0 + b1i + (β3 + b2i)tij + εij, if control.

• Implied marginal mean structure:

. Linear average evolution in each group

. Equal average intercepts

. Different average slopes

• Implied marginal covariance structure (Σi = σ2Ini):

Cov(Yi(t1),Yi(t2)) = 1 t1

+ σ2δ{t1,t2}

= d22t1 t2 + d12(t1 + t2) + d11 + σ2δ{t1,t2}.

• Note that the model implicitly assumes that the variance function is quadraticover time, with positive curvature d22.

• A model which assumes that all variability in subject-specific slopes can beascribed to treatment differences can be obtained by omitting the random slopesb2i from the above model:

Yij = (β0 + b1i) + (β1Li + β2Hi + β3Ci)tij + εij

β0 + b1i + β1tij + εij, if low dose

β0 + b1i + β2tij + εij, if high dose

β0 + b1i + β3tij + εij, if control.

• This is the so-called random-intercepts model

• The same marginal mean structure is obtained as under the model with randomslopes

• Implied marginal covariance structure (Σi = σ2Ini):

Cov(Yi(t1),Yi(t2)) = 1

+ σ2δ{t1,t2}

= d11 + σ2δ{t1,t2}.

• Hence, the implied covariance matrix is compound symmetry:

. constant variance d11 + σ2

. constant correlation ρI = d11/(d11 + σ2) between any two repeatedmeasurements within the same rat

2.7 Example: Bivariate Observations

• Balanced data, two measurements per subject (ni = 2), two models:

Model 1:Random intercepts

+heterogeneous errors

(d) (1 1) +

σ21 0

0 σ22

d + σ21 d

d d+ σ22

Model 2:Uncorrelated intercepts and slopes

+measurement error

d1 + σ2 d1

d1 d1 + d2 + σ2

• Different hierarchical models can produce the same marginal model

• Hence, a good fit of the marginal model cannot be interpreted as evidence for anyof the hierarchical models.

• A satisfactory treatment of the hierarchical model is only possible within aBayesian context.

Chapter 3

Estimation and Inference in the Marginal Model

. ML and REML estimation

. Fitting linear mixed models in SAS

. Negative variance components

. Inference

3.1 ML and REML Estimation

• Recall that the general linear mixed model equals

bi ∼ N (0, D)

εi ∼ N (0,Σi)

independent

• The implied marginal model equals Yi ∼ N (Xiβ, ZiDZ′i + Σi)

• Note that inferences based on the marginal model do not explicitly assume thepresence of random effects representing the natural heterogeneity between subjects

• Notation:

. β: vector of fixed effects (as before)

. α: vector of all variance components in D and Σi

. θ = (β′,α′)′: vector of all parameters in marginal model

• Marginal likelihood function:

LML(θ) =N∏

(2π)−ni/2 |Vi(α)|−

12 exp

2(Yi −Xiβ)′ V −1

i (α) (Yi −Xiβ)

• If α were known, MLE of β equals

β(α) =

i=1X ′iWiXi

i=1X ′iWiyi,

where Wi equals V −1i .

• In most cases, α is not known, and needs to be replaced by an estimate α

• Two frequently used estimation methods for α:

. Maximum likelihood

. Restricted maximum likelihood

3.2 Fitting Linear Mixed Models in SAS

• A model for the prostate data:

ln(PSAij + 1)

= β1Agei + β2Ci + β3Bi + β4Li + β5Mi

+ (β6Agei + β7Ci + β8Bi + β9Li + β10Mi) tij+ (β11Agei + β12Ci + β13Bi + β14Li + β15Mi) t

+ b1i + b2itij + b3it2ij + εij.

• SAS program (Σi = σ2Ini):

proc mixed data=prostate method=reml;

class id group timeclss;

model lnpsa = group age time group*time age*time

time2 group*time2 age*time2 / solution;

random intercept time time2 / type=un subject=id;

• ML and REML estimates for fixed effects:

Effect Parameter MLE (s.e.) REMLE (s.e.)

Age effect β1 0.026 (0.013) 0.027 (0.014)

Intercepts:

Control β2 −1.077 (0.919) −1.098 (0.976)BPH β3 −0.493 (1.026) −0.523 (1.090)

L/R cancer β4 0.314 (0.997) 0.296 (1.059)

Met. cancer β5 1.574 (1.022) 1.549 (1.086)

Age×time effect β6 −0.010 (0.020) −0.011 (0.021)Time effects:

Control β7 0.511 (1.359) 0.568 (1.473)

BPH β8 0.313 (1.511) 0.396 (1.638)L/R cancer β9 −1.072 (1.469) −1.036 (1.593)

Met. cancer β10 −1.657 (1.499) −1.605 (1.626)

Age×time2 effect β11 0.002 (0.008) 0.002 (0.009)

Time2 effects:Control β12 −0.106 (0.549) −0.130 (0.610)

BPH β13 −0.119 (0.604) −0.158 (0.672)

L/R cancer β14 0.350 (0.590) 0.342 (0.656)Met. cancer β15 0.411 (0.598) 0.395 (0.666)

• ML and REML estimates for variance components:

Effect Parameter MLE (s.e.) REMLE (s.e.)

Covariance of bi:

var(b1i) d11 0.398 (0.083) 0.452 (0.098)var(b2i) d22 0.768 (0.187) 0.915 (0.230)

var(b3i) d33 0.103 (0.032) 0.131 (0.041)

cov(b1i, b2i) d12 = d21 −0.443 (0.113) −0.518 (0.136)

cov(b2i, b3i) d23 = d32 −0.273 (0.076) −0.336 (0.095)cov(b3i, b1i) d13 = d31 0.133 (0.043) 0.163 (0.053)

Residual variance:

var(εij) σ2 0.028 (0.002) 0.028 (0.002)

Log-likelihood −1.788 −31.235

• Fitted average profiles at median age at diagnosis:

3.3 Negative Variance Components

• Reconsider the model for the rat data:

Yij = (β0 + b1i) + (β1Li + β2Hi + β3Ci + b2i)tij + εij

• REML estimates obtained from SAS procedure MIXED:

Effect Parameter REMLE (s.e.)Intercept β0 68.606 (0.325)Time effects:

Low dose β1 7.503 (0.228)High dose β2 6.877 (0.231)Control β3 7.319 (0.285)

Covariance of bi:var(b1i) d11 3.369 (1.123)var(b2i) d22 0.000 ( )cov(b1i, b2i) d12 = d21 0.090 (0.381)

Residual variance:var(εij) σ2 1.445 (0.145)

REML log-likelihood −466.173

• This suggests that the REML likelihood could be further increased by allowingnegative estimates for d22

• In SAS, this can be done by adding the option ‘nobound’ to the PROC MIXEDstatement.

• Results: Parameter restrictions for αdii ≥ 0, σ2 ≥ 0 dii ∈ IR, σ2 ∈ IR

Effect Parameter REMLE (s.e.) REMLE (s.e.)

Intercept β0 68.606 (0.325) 68.618 (0.313)

Time effects:

Low dose β1 7.503 (0.228) 7.475 (0.198)

High dose β2 6.877 (0.231) 6.890 (0.198)Control β3 7.319 (0.285) 7.284 (0.254)

Covariance of bi:

var(b1i) d11 3.369 (1.123) 2.921 (1.019)var(b2i) d22 0.000 ( ) −0.287 (0.169)

cov(b1i, b2i) d12 = d21 0.090 (0.381) 0.462 (0.357)

Residual variance:

var(εij) σ2 1.445 (0.145) 1.522 (0.165)

REML log-likelihood −466.173 −465.193

• Note that the REML log-likelihood value has been further increased and anegative estimate for d22 is obtained.

• Brown & Prescott (1999, p. 237) :

• Meaning of negative variance component ?

. Fitted variance function:

Var(Yi(t)) = 1 t

= d22t2 + 2 d12t+ d11 + σ2 = −0.287t2 + 0.924t + 4.443

. The suggested negative curvature in the variance function is supported by thesample variance function:

• This again shows that the hierarchical and marginal models are not equivalent:

. The marginal model allows negative variance components, as long as themarginal covariances Vi = ZiDZ

′i + σ2Ini

are positive definite

. The hierarchical interpretation of the model does not allow negative variancecomponents

3.4 Inference for the Marginal Model

• Estimate for β:

β(α) =

i=1X ′iWiXi

i=1X ′iWiyi

with α replaced by its ML or REML estimate

• Conditional on α, β(α) is multivariate normal with mean β and covariance

Var(β) =

i=1X ′iWiXi

i=1X ′iWivar(Yi)WiXi

i=1X ′iWiXi

• In practice one again replaces α by its ML or REML estimate

• Inference for β:

. Wald tests

. t- and F -tests

. LR tests (not with REML)

• Inference for α:

. Wald tests

. LR tests (even with REML)

. Caution 1: Marginal vs. hierarchical model

. Caution 2: Boundary problems

Chapter 4

Inference for the Random Effects

. Empirical Bayes inference

. Example: Prostate data

. Shrinkage

. Example: Prostate data

4.1 Empirical Bayes Inference

• Random effects bi reflect how the evolution for the ith subject deviates from theexpected evolution Xiβ.

• Estimation of the bi helpful for detecting outlying profiles

• This is only meaningful under the hierarchical model interpretation:

Yi|bi ∼ N (Xiβ + Zibi,Σi) bi ∼ N (0, D)

• Since the bi are random, it is most natural to use Bayesian methods

• Terminology: prior distribution N (0, D) for bi

• Posterior density:

f (bi|yi) ≡ f (bi|Yi = yi) =f (yi|bi) f (bi)

∫f (yi|bi) f (bi) dbi

∝ f (yi|bi) f (bi)

∝ . . .

∝ exp

2(bi −DZ ′iWi(yi −Xiβ))

′Λ−1

i (bi −DZ ′iWi(yi −Xiβ))

for some positive definite matrix Λi.

• Posterior distribution:

bi | yi ∼ N (DZ ′iWi(yi −Xiβ),Λi)

• Posterior mean as estimate for bi:

bi(θ) = E [bi | Yi = yi] =

∫bi f (bi|yi) dbi = DZ ′iWi(α)(yi −Xiβ)

• Parameters in θ are replaced by their ML or REML estimates, obtained fromfitting the marginal model.

• bi =

bi(θ) is called the Empirical Bayes estimate of bi.

• Approximate t- and F -tests to account for the variability introduced by replacingθ by

θ, similar to tests for fixed effects.

4.2 Example: Prostate Data

• We reconsider our linear mixed model:

ln(PSAij + 1)

= β1Agei + β2Ci + β3Bi + β4Li + β5Mi

+ (β6Agei + β7Ci + β8Bi + β9Li + β10Mi) tij+ (β11Agei + β12Ci + β13Bi + β14Li + β15Mi) t

+ b1i + b2itij + b3it2ij + εij.

• In SAS the estimates can be obtained from adding the option ‘solution’ to therandom statement:

random intercept time time2

/ type=un subject=id solution;

ods listing exclude solutionr;

ods output solutionr=out;

• The ODS statements are used to write the EB estimates into a SAS output dataset, and to prevent SAS from printing them in the output window.

• In practice, histograms and scatterplots of certain components of bi are used to

detect model deviations or subjects with ‘exceptional’ evolutions over time

Statistical

Society,

,26–27

201460

• Strong negative correlations in agreement with correlation matrix corresponding tofitted D:

Dcorr =

1.000 −0.803 0.658

−0.803 1.000 −0.968

0.658 −0.968 1.000

• Histograms and scatterplots show outliers

• Subjects #22, #28, #39, and #45, have highest four slopes for time2 andsmallest four slopes for time, i.e., with the strongest (quadratic) growth.

• Subjects #22, #28 and #39 have been further examined and have been shown tobe metastatic cancer cases which were misclassified as local cancer cases.

• Subject #45 is the metastatic cancer case with the strongest growth

Part II

Marginal Models for Non-Gaussian Longitudinal Data

Chapter 5

The Toenail Data

• Toenail Dermatophyte Onychomycosis: Common toenail infection, difficult totreat, affecting more than 2% of population.

• Classical treatments with antifungal compounds need to be administered until thewhole nail has grown out healthy.

• New compounds have been developed which reduce treatment to 3 months

• Randomized, double-blind, parallel group, multicenter study for the comparison oftwo such new compounds (A and B) for oral treatment.

• Research question:

Severity relative to treatment of TDO ?

• 2× 189 patients randomized, 36 centers

• 48 weeks of total follow up (12 months)

• 12 weeks of treatment (3 months)

• measurements at months 0, 1, 2, 3, 6, 9, 12.

• Frequencies at each visit (both treatments):

Chapter 6

The Analgesic Trial

• single-arm trial with 530 patients recruited (491 selected for analysis)

• analgesic treatment for pain caused by chronic nonmalignant disease

• treatment was to be administered for 12 months

• we will focus on Global Satisfaction Assessment (GSA)

• GSA scale goes from 1=very good to 5=very bad

• GSA was rated by each subject 4 times during the trial, at months 3, 6, 9, and 12.

• Research questions:

. Evolution over time

. Relation with baseline covariates: age, sex, duration of the pain, type of pain,disease progression, Pain Control Assessment (PCA), . . .

. Investigation of dropout

• Frequencies:

GSA Month 3 Month 6 Month 9 Month 12

1 55 14.3% 38 12.6% 40 17.6% 30 13.5%

2 112 29.1% 84 27.8% 67 29.5% 66 29.6%

3 151 39.2% 115 38.1% 76 33.5% 97 43.5%

4 52 13.5% 51 16.9% 33 14.5% 27 12.1%

5 15 3.9% 14 4.6% 11 4.9% 3 1.4%

Tot 385 302 227 223

• Missingness:

Measurement occasion

Month 3 Month 6 Month 9 Month 12 Number %

Completers

O O O O 163 41.2

Dropouts

O O O M 51 12.91

O O M M 51 12.91

O M M M 63 15.95

Non-monotone missingness

O O M O 30 7.59

O M O O 7 1.77

O M O M 2 0.51

O M M O 18 4.56

M O O O 2 0.51

M O O M 1 0.25

M O M O 1 0.25

M O M M 3 0.76

Chapter 7

Generalized Linear Models

. The model

. Maximum likelihood estimation

. Examples

. McCullagh and Nelder (1989)

7.1 The Generalized Linear Model

• Suppose a sample Y1, . . . , YN of independent observations is available

• All Yi have densities f (yi|θi, φ) which belong to the exponential family:

f (y|θi, φ) = exp{φ−1[yθi − ψ(θi)] + c(y, φ)

• θi the natural parameter

• Linear predictor: θi = xi′β

• φ is the scale parameter (overdispersion parameter)

• ψ(·) is a function, generating mean and variance:

E(Y ) = ψ′(θ)

Var(Y ) = φψ′′(θ)

• Note that, in general, the mean µ and the variance are related:

Var(Y ) = φψ′′[

ψ′−1(µ)

= φv(µ)

• The function v(µ) is called the variance function.

• The function ψ′−1 which expresses θ as function of µ is called the link function.

• ψ′ is the inverse link function

7.2 Examples

7.2.1 The Normal Model

• Model:

Y ∼ N (µ, σ2)

• Density function:

f (y|θ, φ) =1√

2πσ2exp

σ2(y − µ)2

yµ− µ2

ln(2πσ2)

2− y2

• Exponential family:

. θ = µ

. φ = σ2

. ψ(θ) = θ2/2

. c(y, φ) = ln(2πφ)2− y2

• Mean and variance function:

. µ = θ

. v(µ) = 1

• Note that, under this normal model, the mean and variance are not related:

φv(µ) = σ2

• The link function is here the identity function: θ = µ

7.2.2 The Bernoulli Model

• Model:Y ∼ Bernoulli(π)

• Density function:

f (y|θ, φ) = πy(1− π)1−y

= exp {y ln π + (1− y) ln(1− π)}

1− π

+ ln(1− π)

• Exponential family:

. θ = ln(

π1−π

. φ = 1

. ψ(θ) = ln(1− π) = ln(1 + exp(θ))

. c(y, φ) = 0

• Mean and variance function:

. µ = exp θ1+exp θ = π

. v(µ) = exp θ(1+exp θ)2

= π(1− π)

• Note that, under this model, the mean and variance are related:

φv(µ) = µ(1− µ)

• The link function here is the logit link: θ = ln(

µ1−µ

7.3 Generalized Linear Models (GLM)

• Suppose a sample Y1, . . . , YN of independent observations is available

• All Yi have densities f (yi|θi, φ) which belong to the exponential family

• In GLM’s, it is believed that the differences between the θi can be explainedthrough a linear function of known covariates:

θi = xi′β

• xi is a vector of p known covariates

• β is the corresponding vector of unknown regression parameters, to be estimatedfrom the data.

7.4 Maximum Likelihood Estimation

• Log-likelihood:

`(β, φ) =1

i[yiθi − ψ(θi)] +

ic(yi, φ)

• The score equations:

S(β) =∑

∂µi

∂βv−1

i (yi − µi) = 0

• Note that the estimation of β depends on the density only through the means µi

and the variance functions vi = v(µi).

• The score equations need to be solved numerically:

. iterative (re-)weighted least squares

. Newton-Raphson

. Fisher scoring

• Inference for β is based on classical maximum likelihood theory:

. asymptotic Wald tests

. likelihood ratio tests

. score tests

7.5 Illustration: The Analgesic Trial

• Early dropout (did the subject drop out after the first or the second visit) ?

• Binary response

• PROC GENMOD can fit GLMs in general

• PROC LOGISTIC can fit models for binary (and ordered) responses

• SAS code for logit link:

proc genmod data=earlydrp;

model earlydrp = pca0 weight psychiat physfct / dist=b;

proc logistic data=earlydrp descending;

model earlydrp = pca0 weight psychiat physfct;

• SAS code for probit link:

proc genmod data=earlydrp;

model earlydrp = pca0 weight psychiat physfct / dist=b link=probit;

proc logistic data=earlydrp descending;

model earlydrp = pca0 weight psychiat physfct / link=probit;

• Selected output:

Analysis Of Parameter Estimates

Standard Wald 95% Chi-

Parameter DF Estimate Error Confidence Limits Square Pr > ChiSq

Intercept 1 -1.0673 0.7328 -2.5037 0.3690 2.12 0.1453

PCA0 1 0.3981 0.1343 0.1349 0.6614 8.79 0.0030

WEIGHT 1 -0.0211 0.0072 -0.0353 -0.0070 8.55 0.0034

PSYCHIAT 1 0.7169 0.2871 0.1541 1.2796 6.23 0.0125

PHYSFCT 1 0.0121 0.0050 0.0024 0.0219 5.97 0.0145

Scale 0 1.0000 0.0000 1.0000 1.0000

NOTE: The scale parameter was held fixed.

Chapter 8

Parametric Modeling Families

. Continuous outcomes

. Longitudinal generalized linear models

. Notation

8.1 Continuous Outcomes

•Marginal Models:

E(Yij|xij) = x′ijβ

• Random-Effects Models:

E(Yij|bi,xij) = x′ijβ + z′ijbi

• Transition Models:

E(Yij|Yi,j−1, . . . , Yi1,xij) = x′ijβ + αYi,j−1

8.2 Longitudinal Generalized Linear Models

• Normal case: easy transfer between models

• Also non-normal data can be measured repeatedly (over time)

• Lack of key distribution such as the normal [=⇒]

. A lot of modeling options

. Introduction of non-linearity

. No easy transfer between model families

cross-sectional longitudinal

normal outcome linear model LMM

non-normal outcome GLM ?

8.3 Marginal Models

• Fully specified models & likelihood estimation

. E.g., multivariate probit model, Bahadur model, odds ratio model

. Specification and/or fitting can be cumbersome

• Non-likelihood alternatives

. Historically: Empirical generalized least squares

(EGLS; Koch et al 1977; SAS CATMOD)

. Generalized estimating equations

. Pseudo-likelihood

Chapter 9

Generalized Estimating Equations

. General idea

. Asymptotic properties

. Working correlation

. Special case and application

. SAS code and output

9.1 General Idea

• Univariate GLM, score function of the form (scalar Yi):

S(β) =N∑

∂µi

∂βv−1

i (yi − µi) = 0 with vi = Var(Yi)

• In longitudinal setting: Y = (Y 1, . . . ,Y N):

S(β) =∑

∂µij

∂βv−1

ij (yij − µij) =N∑

i=1D′i [Vi(α)]−1 (yi − µi) = 0

. Di is an ni × p matrix with (i, j)th elements∂µij

. yi and µi are ni-vectors with elements yij and µij

. Is Vi ni × ni diagonal or more complex?

• Vi = Var(Y i) is more complex since it involves a set of nuisance parameters α,determining the covariance structure of Y i:

Vi(β,α) = φA1/2i (β)Ri(α)A

1/2i (β)

in which

A1/2i (β) =

√vi1(µi1(β)) . . . 0

... . . . ...

0 . . .√vini

(µini(β))

and Ri(α) is the correlation matrix of Yi, parameterized by α.

• Same form as for full likelihood procedure, but we restrict specification to the firstmoment only

• Liang and Zeger (1986)

9.2 Large Sample Properties

As N →∞ √N(β − β) ∼ N (0, I−1

I0 =N∑

i=1D′i[Vi(α)]−1Di

• (Unrealistic) Conditions:

. α is known

. the parametric form for Vi(α) is known

• This is the naive≡purely model based variance estimator

• Solution: working correlation matrix

9.3 Unknown Covariance Structure

Keep the score equations

S(β) =N∑

i=1[Di]

′ [Vi(α)]−1 (yi − µi) = 0

• suppose Vi(.) is not the true variance of Y i but only a plausible guess, a so-calledworking correlation matrix

• specify correlations and not covariances, because the variances follow from themean structure

• the score equations are solved as before

• The asymptotic normality results change to

√N(β − β) ∼ N (0, I−1

0 I1I−10 )

I0 =N∑

i=1D′i[Vi(α)]−1Di

I1 =N∑

i=1D′i[Vi(α)]−1Var(Y i)[Vi(α)]−1Di.

• This is the robust≡empirically corrected≡ sandwich variance estimator

. I0 is the bread

. I1 is the filling (ham or cheese)

• Correct guess =⇒ likelihood variance

• The estimators β are consistent even if the working correlation matrix is incorrect

• An estimate is found by replacing the unknown variance matrix Var(Y i) by

(Y i − µi)(Y i − µi)′.

• Even if this estimator is bad for Var(Y i) it leads to a good estimate of I1,provided that:

. replication in the data is sufficiently large

. same model for µi is fitted to groups of subjects

. observation times do not vary too much between subjects

• A bad choice of working correlation matrix can affect the efficiency of β

• Care needed with incomplete data (see Part IV)

9.4 The Working Correlation Matrix

Vi = Vi(β,α, φ) = φA1/2i (β)Ri(α)A

1/2i (β)

• Variance function: Ai is (ni × ni) diagonal with elements v(µij), the known GLMvariance function.

• Working correlation: Ri(α) possibly depends on a different set of parameters α.

• Overdispersion parameter: φ, assumed 1 or estimated from the data.

• The unknown quantities are expressed in terms of the Pearson residuals

eij =yij − µij√v(µij)

Note that eij depends on β.

9.5 Estimation of Working Correlation

Liang and Zeger (1986) proposed moment-based estimates for the working correlation.

Corr(Yij, Yik) Estimate

Independence 0 —

Exchangeable α α = 1N

∑Ni=1

1ni(ni−1)

∑j 6=k eijeik

AR(1) α α = 1N

∑Ni=1

1ni−1

∑j≤ni−1 eijei,j+1

Unstructured αjk αjk = 1N

∑Ni=1 eijeik

Dispersion parameter:

j=1e2ij.

9.6 Fitting GEE

The standard procedure, implemented in the SAS procedure GENMOD.

1. Compute initial estimates for β, using a univariate GLM (i.e., assumingindependence).

2. . Compute Pearson residuals eij.

. Compute estimates for α and φ.

. Compute Ri(α) and Vi(β,α) = φA1/2i (β)Ri(α)A

1/2i (β).

3. Update estimate for β:

β(t+1) = β(t) −

i=1D′iV

−1i Di

i=1D′iV

−1i (yi − µi)

4. Iterate 2.–3. until convergence.

Estimates of precision by means of I−10 and/or I−1

0 I1I−10 .

9.7 Special Case: Linear Mixed Models

• Estimate for β:

β(α) =

i=1X ′iWiXi

i=1X ′iWiYi

with α replaced by its ML or REML estimate

• Conditional on α, β has mean

E[β(α)

i=1X ′iWiXi

i=1X ′iWiXiβ = β

provided that E(Yi) = Xiβ

• Hence, in order for β to be unbiased, it is sufficient that the mean of the response

is correctly specified.

• Conditional on α, β has covariance

Var(β) =

i=1X ′iWiXi

i=1X ′iWiVar(Yi)WiXi

i=1X ′iWiXi

• Note that this model-based version assumes that the covariance matrixVar(Yi) is correctly modelled as Vi = ZiDZ

′i + Σi.

• An empirically corrected version is:

Var(β) =

i=1X ′iWiXi

︸︷︷︸

↓BREAD

i=1X ′iWiVar(Yi)WiXi

︸︷︷︸

↓MEAT

i=1X ′iWiXi

︸︷︷︸

↓BREAD

Chapter 10

A Family of GEE Methods

. Classical approach

. Prentice’s two sets of GEE

. Linearization-based version

. GEE2

. Alternating logistic regressions

10.1 Prentice’s GEE

i=1D′iV

−1i (Y i − µi) = 0,

i=1E′iW

−1i (Z i − δi) = 0

Zijk =(Yij − µij)(Yik − µik)

√µij(1− µij)µik(1− µik)

, δijk = E(Zijk)

The joint asymptotic distribution of√N (β − β) and

√N (α−α) normal with

variance-covariance matrix consistently estimated by

Λ11 Λ12

Λ21 Λ22

A B′

D′iV−1

E ′iW−1

∂Z i

D′iV−1

E ′iW−1

Λ11 =N∑

D′iV−1

i Cov(Y i)V−1

Λ12 =N∑

D′iV−1

i Cov(Y i,Zi)W−1

Λ21 = Λ12,

Λ22 =N∑

E ′iW−1

i Cov(Zi)W−1

Statistic Estimator

Var(Y i) (Y i − µi)(Y i − µi)′

Cov(Y i,Zi) (Y i − µi)(Z i − δi)′

Var(Z i) (Z i − δi)(Z i − δi)′

10.2 GEE Based on Linearization

10.2.1 Model formulation

• Previous version of GEE are formulated directly in terms of binary outcomes

• This approach is based on a linearization:

yi = µi + εi

ηi = g(µi), ηi = X iβ, Var(yi) = Var(εi) = Σi.

• ηi is a vector of linear predictors,

• g(.) is the (vector) link function.

10.2.2 Estimation (Nelder and Wedderburn 1972)

• Solve iteratively:N∑

i=1X ′iWiXiβ =

i=1Wiy

∗i ,

Wi = F ′iΣ−1i Fi, y∗i = ηi + (yi − µi)F

−1i ,

Fi =∂µi

∂ηi

, Σi = Var(ε), µi = E(yi).

• Remarks:

. y∗i is called ‘working variable’ or ‘pseudo data’.

. Basis for SAS macro and procedure GLIMMIX

. For linear models, Di = Iniand standard linear regression follows.

10.2.3 The Variance Structure

Σi = φA1/2i (β)Ri(α)A

1/2i (β)

• φ is a scale (overdispersion) parameter,

• Ai = v(µi), expressing the mean-variance relation (this is a function of β),

• Ri(α) describes the correlation structure:

. If independence is assumed then Ri(α) = Ini.

. Other structures, such as compound symmetry, AR(1),. . . can be assumed aswell.

10.3 GEE2

•Model:

. Marginal mean structure

. Pairwise association:

∗ Odds ratios

∗ Correlations

•Working assumptions: Third and fourth moments

• Estimation:

. Second-order estimating equations

. Likelihood (assuming 3rd and 4th moments are correctly specified)

10.4 Alternating Logistic Regression

• Diggle, Heagerty, Liang, and Zeger (2002) and Molenberghs and Verbeke (2005)

• When marginal odds ratios are used to model association, α can be estimatedusing ALR, which is

. almost as efficient as GEE2

. almost as easy (computationally) than GEE1

• µijk as before and αijk = ln(ψijk) the marginal log odds ratio:

logit Pr(Yij = 1|xij) = xijβ

logit Pr(Yij = 1|Yik = yik) = αijkyik + ln

µij − µijk

1− µij − µik + µijk

• αijk can be modelled in terms of predictors

• the second term is treated as an offset

• the estimating equations for β and α are solved in turn, and the ‘alternating’between both sets is repeated until convergence.

• this is needed because the offset clearly depends on β.

10.5 Application to the Toenail Data

10.5.1 The model

• Consider the model:

Yij ∼ Bernoulli(µij), log

1− µij

= β0 + β1Ti + β2tij + β3Titij

• Yij: severe infection (yes/no) at occasion j for patient i

• tij: measurement time for occasion j

• Ti: treatment group

10.5.2 Standard GEE

• SAS Code:

proc genmod data=test descending;

class idnum timeclss;

model onyresp = treatn time treatn*time

/ dist=binomial;

repeated subject=idnum / withinsubject=timeclss

type=exch covb corrw modelse;

• SAS statements:

. The REPEATED statements defines the GEE character of the model.

. ‘type=’: working correlation specification (UN, AR(1), EXCH, IND,. . . )

. ‘modelse’: model-based s.e.’s on top of default empirically corrected s.e.’s

. ‘corrw’: printout of working correlation matrix

. ‘withinsubject=’: specification of the ordering within subjects

. Regression parameters:

Analysis Of Initial Parameter Estimates

Standard Wald 95% Chi-

Parameter DF Estimate Error Confidence Limits Square

Intercept 1 -0.5571 0.1090 -0.7708 -0.3433 26.10

treatn 1 0.0240 0.1565 -0.2827 0.3307 0.02

time 1 -0.1769 0.0246 -0.2251 -0.1288 51.91

treatn*time 1 -0.0783 0.0394 -0.1556 -0.0010 3.95

Scale 0 1.0000 0.0000 1.0000 1.0000

. Estimates from fitting the model, ignoring the correlation structure, i.e.,from fitting a classical GLM to the data, using proc GENMOD.

. The reported log-likelihood also corresponds to this model, and thereforeshould not be interpreted.

. The reported estimates are used as starting values in the iterative estimationprocedure for fitting the GEE’s.

Analysis Of GEE Parameter Estimates

Empirical Standard Error Estimates

Standard 95% Confidence

Parameter Estimate Error Limits Z Pr > |Z|

Intercept -0.5840 0.1734 -0.9238 -0.2441 -3.37 0.0008

treatn 0.0120 0.2613 -0.5001 0.5241 0.05 0.9633

time -0.1770 0.0311 -0.2380 -0.1161 -5.69 <.0001

treatn*time -0.0886 0.0571 -0.2006 0.0233 -1.55 0.1208

Model-Based Standard Error Estimates

Intercept -0.5840 0.1344 -0.8475 -0.3204 -4.34 <.0001

treatn 0.0120 0.1866 -0.3537 0.3777 0.06 0.9486

time -0.1770 0.0209 -0.2180 -0.1361 -8.47 <.0001

treatn*time -0.0886 0.0362 -0.1596 -0.0177 -2.45 0.0143

. The working correlation:

Exchangeable Working Correlation

Correlation 0.420259237

10.5.3 Alternating Logistic Regression

• ‘type=exch’ −→ ‘logor=exch’

• Note that α now is a genuine parameter

Empirical Standard Error Estimates

Intercept -0.5244 0.1686 -0.8548 -0.1940 -3.11 0.0019

treatn 0.0168 0.2432 -0.4599 0.4935 0.07 0.9448

time -0.1781 0.0296 -0.2361 -0.1200 -6.01 <.0001

treatn*time -0.0837 0.0520 -0.1856 0.0182 -1.61 0.1076

Alpha1 3.2218 0.2908 2.6519 3.7917 11.08 <.0001

Model-Based Standard Error Estimates

Intercept -0.5244 0.1567 -0.8315 -0.2173 -3.35 0.0008

treatn 0.0168 0.2220 -0.4182 0.4519 0.08 0.9395

time -0.1781 0.0233 -0.2238 -0.1323 -7.63 <.0001

treatn*time -0.0837 0.0392 -0.1606 -0.0068 -2.13 0.0329

10.5.4 Linearization Based Method

• GLIMMIX macro:

%glimmix(data=test, procopt=%str(method=ml empirical),

stmts=%str(

class idnum timeclss;

model onyresp = treatn time treatn*time / solution;

repeated timeclss / subject=idnum type=cs rcorr;

error=binomial,

link=logit);

• GLIMMIX procedure:

proc glimmix data=test method=RSPL empirical;

class idnum;

model onyresp (event=’1’) = treatn time treatn*time

/ dist=binary solution;

random _residual_ / subject=idnum type=cs;

• Both produce the same results

• The GLIMMIX macro is a MIXED core, with GLM-type surrounding statements

• The GLIMMIX procedure does not call MIXED, it has its own engine

• PROC GLIMMIX combines elements of MIXED and of GENMOD

• RANDOM residual is the PROC GLIMMIX way to specify residual correlation

10.5.5 Results of Models Fitted to Toenail Data

Effect Par. IND EXCH UN

Int. β0 -0.557(0.109;0.171) -0.584(0.134;0.173) -0.720(0.166;0.173)

Ti β1 0.024(0.157;0.251) 0.012(0.187;0.261) 0.072(0.235;0.246)

tij β2 -0.177(0.025;0.030) -0.177(0.021;0.031) -0.141(0.028;0.029)

Ti · tij β3 -0.078(0.039;0.055) -0.089(0.036;0.057) -0.114(0.047;0.052)

Int. β0 -0.524(0.157;0.169)

Ti β1 0.017(0.222;0.243)

tij β2 -0.178(0.023;0.030)

Ti · tij β3 -0.084(0.039;0.052)

Ass. α 3.222( ;0.291)

Linearization based method

Int. β0 -0.557(0.112;0.171) -0.585(0.142;0.174) -0.630(0.171;0.172)

Ti β1 0.024(0.160;0.251) 0.011(0.196;0.262) 0.036(0.242;0.242)

tij β2 -0.177(0.025;0.030) -0.177(0.022;0.031) -0.204(0.038;0.034)

Ti · tij β3 -0.078(0.040:0.055) -0.089(0.038;0.057) -0.106(0.058;0.058)

estimate (model-based s.e.; empirical s.e.)

10.5.6 Discussion

• GEE1: All empirical standard errors are correct, but the efficiency is higher for themore complex working correlation structure, as seen in p-values for Ti · tij effect:

Structure p-value

IND 0.1515

EXCH 0.1208

UN 0.0275

Thus, opting for reasonably adequate correlation assumptions still pays off, inspite of the fact that all are consistent and asymptotically normal

• Similar conclusions for linearization-based method

• Model-based s.e. and empirically corrected s.e. in reasonable agreement for UN

• Typically, the model-based standard errors are much too small as they are basedon the assumption that all observations in the data set are independent, herebyoverestimating the amount of available information, hence also overestimating theprecision of the estimates.

• ALR: similar inferences but now also α part of the inferences

Part III

Generalized Linear Mixed Models for Non-GaussianLongitudinal Data

Chapter 11

Generalized Linear Mixed Models (GLMM)

. Introduction: LMM Revisited

. Generalized Linear Mixed Models (GLMM)

. Fitting Algorithms

. Example

11.1 Introduction: LMM Revisited

• We re-consider the linear mixed model:

Yi|bi ∼ N (Xiβ + Zibi,Σi), bi ∼ N (0,D)

• The implied marginal model equals Yi ∼ N (Xiβ, ZiDZ′i + Σi)

• Hence, even under conditional independence, i.e., all Σi equal to σ2Ini, a marginal

association structure is implied through the random effects.

• The same ideas can now be applied in the context of GLM’s to model associationbetween discrete repeated measures.

11.2 Generalized Linear Mixed Models (GLMM)

• Given a vector bi of random effects for cluster i, it is assumed that all responsesYij are independent, with density

f (yij|θij, φ) = exp{φ−1[yijθij − ψ(θij)] + c(yij, φ)

• θij is now modelled as θij = xij′β + zij

• As before, it is assumed that bi ∼ N (0, D)

• Let fij(yij|bi,β, φ) denote the conditional density of Yij given bi, the conditionaldensity of Yi equals

fi(yi|bi,β, φ) =ni∏

j=1fij(yij|bi,β, φ)

• The marginal distribution of Yi is given by

fi(yi|β, D, φ) =∫fi(yi|bi,β, φ) f (bi|D) dbi

=∫ ni∏

j=1fij(yij|bi,β, φ) f (bi|D) dbi

where f (bi|D) is the density of the N (0,D) distribution.

• The likelihood function for β, D, and φ now equals

L(β, D, φ) =N∏

i=1fi(yi|β, D, φ)

∫ ni∏

• Under the normal linear model, the integral can be worked out analytically.

• In general, approximations are required:

. Approximation of integrand

. Approximation of data

. Approximation of integral

• Predictions of random effects can be based on the posterior distribution

f (bi|Yi = yi)

• ‘Empirical Bayes (EB) estimate’:Posterior mode, with unknown parameters replaced by their MLE

11.3 Laplace Approximation of Integrand

• Integrals in L(β,D, φ) can be written in the form I =∫eQ(b)db

• Second-order Taylor expansion of Q(b) around the mode yields

Q(b) ≈ Q(b) +1

2(b−

b)′Q′′(b)(b − b),

• Quadratic term leads to re-scaled normal density. Hence,

I ≈ (2π)q/2∣∣∣∣−Q′′(b)

∣∣∣∣−1/2

eQ(b).

• Exact approximation in case of normal kernels

• Good approximation in case of many repeated measures per subject

11.4 Approximation of Data

11.4.1 General Idea

• Re-write GLMM as:

Yij = µij + εij = h(x′ijβ + z′ijbi) + εij

with variance for errors equal to Var(Yij|bi) = φv(µij)

• Linear Taylor expansion for µij:

. Penalized quasi-likelihood (PQL): Around current β and

. Marginal quasi-likelihood (MQL): Around current β and bi = 0

11.4.2 Penalized quasi-likelihood (PQL)

• Linear Taylor expansion around current β and

Yij ≈ h(x′ijβ + z′ijbi) + h′(x′ijβ + z′ijbi)x′ij(β − β) + h′(x′ijβ + z′ijbi)z

′ij(bi − bi) + εij

≈ µij + v(µij)x′ij(β − β) + v(µij)z

′ij(bi − bi) + εij

• In vector notation: Yi ≈ µi + ViXi(β − β) + ViZi(bi −

bi) + εi

• Re-ordering terms yields:

Yi∗ ≡ V −1

i (Yi − µi) +Xiβ + Zi

bi ≈ Xiβ + Zibi + ε∗

• Model fitting by iterating between updating the pseudo responses Yi∗ and fitting

the above linear mixed model to them.

11.4.3 Marginal quasi-likelihood (MQL)

• Linear Taylor expansion around current β and bi = 0:

Yij ≈ h(x′ijβ) + h′(x′ijβ)x′ij(β − β) + h′(x′ijβ)z′ijbi + εij

≈ µij + v(µij)x′ij(β − β) + v(µij)z

′ijbi + εij

• In vector notation: Yi ≈ µi + ViXi(β − β) + ViZibi + εi

• Re-ordering terms yields:

Yi∗ ≡ V −1

i (Yi − µi) +Xiβ ≈ Xiβ + Zibi + ε∗

• Model fitting by iterating between updating the pseudo responses Yi∗ and fitting

the above linear mixed model to them.

11.4.4 PQL versus MQL

• MQL only performs reasonably well if random-effects variance is (very) small

• Both perform bad for binary outcomes with few repeated measurements per cluster

• With increasing number of measurements per subject:

. MQL remains biased

. PQL consistent

• Improvements possible with higher-order Taylor expansions

11.5 Approximation of Integral

• The likelihood contribution of every subject is of the form∫f (z)φ(z)dz

where φ(z) is the density of the (multivariate) normal distribution

• Gaussian quadrature methods replace the integral by a weighted sum:

∫f (z)φ(z)dz ≈

q=1wqf (zq)

• Q is the order of the approximation. The higher Q the more accurate theapproximation will be

• The nodes (or quadrature points) zq are solutions to the Qth order Hermitepolynomial

• The wq are well-chosen weights

• The nodes zq and weights wq are reported in tables. Alternatively, an algorithm isavailable for calculating all zq and wq for any value Q.

• With Gaussian quadrature, the nodes and weights are fixed, independent off (z)φ(z).

• With adaptive Gaussian quadrature, the nodes and weights are adapted tothe ‘support’ of f (z)φ(z).

• Graphically (Q = 10):

• Typically, adaptive Gaussian quadrature needs (much) less quadrature points thanclassical Gaussian quadrature.

• On the other hand, adaptive Gaussian quadrature is much more time consuming.

• Adaptive Gaussian quadrature of order one is equivalent to Laplace transformation.

11.6 Example: Toenail Data

• Yij is binary severity indicator for subject i at visit j.

• Model:

Yij|bi ∼ Bernoulli(πij), log

1− πij

= β0 + bi + β1Ti + β2tij + β3Titij

• Notation:

. Ti: treatment indicator for subject i

. tij: time point at which jth measurement is taken for ith subject

• Adaptive as well as non-adaptive Gaussian quadrature, for various Q.

• Results: Gaussian quadrature

Q = 3 Q = 5 Q = 10 Q = 20 Q = 50

β0 -1.52 (0.31) -2.49 (0.39) -0.99 (0.32) -1.54 (0.69) -1.65 (0.43)

β1 -0.39 (0.38) 0.19 (0.36) 0.47 (0.36) -0.43 (0.80) -0.09 (0.57)

β2 -0.32 (0.03) -0.38 (0.04) -0.38 (0.05) -0.40 (0.05) -0.40 (0.05)

β3 -0.09 (0.05) -0.12 (0.07) -0.15 (0.07) -0.14 (0.07) -0.16 (0.07)

σ 2.26 (0.12) 3.09 (0.21) 4.53 (0.39) 3.86 (0.33) 4.04 (0.39)

−2` 1344.1 1259.6 1254.4 1249.6 1247.7

Adaptive Gaussian quadrature

Q = 3 Q = 5 Q = 10 Q = 20 Q = 50

β0 -2.05 (0.59) -1.47 (0.40) -1.65 (0.45) -1.63 (0.43) -1.63 (0.44)

β1 -0.16 (0.64) -0.09 (0.54) -0.12 (0.59) -0.11 (0.59) -0.11 (0.59)

β2 -0.42 (0.05) -0.40 (0.04) -0.41 (0.05) -0.40 (0.05) -0.40 (0.05)

β3 -0.17 (0.07) -0.16 (0.07) -0.16 (0.07) -0.16 (0.07) -0.16 (0.07)

σ 4.51 (0.62) 3.70 (0.34) 4.07 (0.43) 4.01 (0.38) 4.02 (0.38)

−2` 1259.1 1257.1 1248.2 1247.8 1247.8

• Conclusions:

. (Log-)likelihoods are not comparable

. Different Q can lead to considerable differences in estimates and standarderrors

. For example, using non-adaptive quadrature, with Q = 3, we found nodifference in time effect between both treatment groups(t = −0.09/0.05, p = 0.0833).

. Using adaptive quadrature, with Q = 50, we find a significant interactionbetween the time effect and the treatment (t = −0.16/0.07, p = 0.0255).

. Assuming that Q = 50 is sufficient, the ‘final’ results are well approximatedwith smaller Q under adaptive quadrature, but not under non-adaptivequadrature.

• Comparison of fitting algorithms:

. Adaptive Gaussian Quadrature, Q = 50

. MQL and PQL

• Summary of results:

Parameter QUAD PQL MQL

Intercept group A −1.63 (0.44) −0.72 (0.24) −0.56 (0.17)

Intercept group B −1.75 (0.45) −0.72 (0.24) −0.53 (0.17)

Slope group A −0.40 (0.05) −0.29 (0.03) −0.17 (0.02)

Slope group B −0.57 (0.06) −0.40 (0.04) −0.26 (0.03)

Var. random intercepts (τ 2) 15.99 (3.02) 4.71 (0.60) 2.49 (0.29)

• Severe differences between QUAD (gold standard ?) and MQL/PQL.

• MQL/PQL may yield (very) biased results, especially for binary data.

Chapter 12

Fitting GLMM’s in SAS

. Overview

. Example: Toenail data

12.1 Overview

• GLIMMIX: Laplace, MQL, PQL, adaptive quadrature

• NLMIXED: Adaptive and non-adaptive quadrature−→ not discussed here

12.2 Example: Toenail data

• Re-consider logistic model with random intercepts for toenail data

• SAS code (PQL):

proc glimmix data=test method=RSPL ;

class idnum;

model onyresp (event=’1’) = treatn time treatn*time

/ dist=binary solution;

random intercept / subject=idnum;

• MQL obtained with option ‘method=RMPL’

• Laplace obtained with option ‘method=LAPLACE’

• Adaptive quadrature with option ‘method=QUAD(qpoints=5)’

• Inclusion of random slopes:

random intercept time / subject=idnum type=un;

Part IV

Marginal Versus Random-effects Models

Chapter 13

Marginal Versus Random-effects Models

. Interpretation of GLMM parameters

. Marginalization of GLMM

. Example: Toenail data

13.1 Interpretation of GLMM Parameters: Toenail Data

• We compare our GLMM results for the toenail data with those from fitting GEE’s(unstructured working correlation):

GLMM GEE

Parameter Estimate (s.e.) Estimate (s.e.)

Intercept group A −1.6308 (0.4356) −0.7219 (0.1656)

Intercept group B −1.7454 (0.4478) −0.6493 (0.1671)

Slope group A −0.4043 (0.0460) −0.1409 (0.0277)

Slope group B −0.5657 (0.0601) −0.2548 (0.0380)

• The strong differences can be explained as follows:

. Consider the following GLMM:

Yij|bi ∼ Bernoulli(πij), log

1− πij

= β0 + bi + β1tij

. The conditional means E(Yij|bi), as functions of tij, are given by

E(Yij|bi)

=exp(β0 + bi + β1tij)

1 + exp(β0 + bi + β1tij)

. The marginal average evolution is now obtained from averaging over therandom effects:

E(Yij) = E[E(Yij|bi)] = E

exp(β0 + bi + β1tij)

1 + exp(β0 + bi + β1tij)

6= exp(β0 + β1tij)

1 + exp(β0 + β1tij)

• Hence, the parameter vector β in the GEE model needs to be interpretedcompletely different from the parameter vector β in the GLMM:

. GEE: marginal interpretation

. GLMM: conditional interpretation, conditionally upon level of random effects

• In general, the model for the marginal average is not of the same parametric formas the conditional average in the GLMM.

• For logistic mixed models, with normally distributed random random intercepts, itcan be shown that the marginal model can be well approximated by again alogistic model, but with parameters approximately satisfying

M=√c2σ2 + 1 > 1, σ2 = variance random intercepts

c = 16√

3/(15π)

• For the toenail application, σ was estimated as 4.0164, such that the ratio equals√c2σ2 + 1 = 2.5649.

• The ratio’s between the GLMM and GEE estimates are:

GLMM GEE

Parameter Estimate (s.e.) Estimate (s.e.) Ratio

Intercept group A −1.6308 (0.4356) −0.7219 (0.1656) 2.2590

Intercept group B −1.7454 (0.4478) −0.6493 (0.1671) 2.6881

Slope group A −0.4043 (0.0460) −0.1409 (0.0277) 2.8694

Slope group B −0.5657 (0.0601) −0.2548 (0.0380) 2.2202

• Note that this problem does not occur in linear mixed models:

. Conditional mean: E(Yi|bi) = Xiβ + Zibi

. Specifically: E(Yi|bi = 0) = Xiβ

. Marginal mean: E(Yi) = Xiβ

• The problem arises from the fact that, in general,

E[g(Y )] 6= g[E(Y )]

• So, whenever the random effects enter the conditional mean in a non-linear way,the regression parameters in the marginal model need to be interpreted differentlyfrom the regression parameters in the mixed model.

• In practice, the marginal mean can be derived from the GLMM output byintegrating out the random effects.

• This can be done numerically via Gaussian quadrature, or based on samplingmethods.

13.2 Marginalization of GLMM: Toenail Data

• As an example, we plot the average evolutions based on the GLMM outputobtained in the toenail example:

P (Yij = 1)

exp(−1.6308 + bi − 0.4043tij)

1 + exp(−1.6308 + bi − 0.4043tij)

exp(−1.7454 + bi − 0.5657tij)

1 + exp(−1.7454 + bi − 0.5657tij)

• Average evolutions obtained from the GEE analyses:

P (Yij = 1)

exp(−0.7219− 0.1409tij)

1 + exp(−0.7219− 0.1409tij)

exp(−0.6493− 0.2548tij)

1 + exp(−0.6493− 0.2548tij)

• In a GLMM context, rather than plotting the marginal averages, one can also plotthe profile for an ‘average’ subject, i.e., a subject with random effect bi = 0:

P (Yij = 1|bi = 0)

exp(−1.6308− 0.4043tij)

1 + exp(−1.6308− 0.4043tij)

exp(−1.7454− 0.5657tij)

1 + exp(−1.7454− 0.5657tij)

13.3 Example: Toenail Data Revisited

• Overview of all analyses on toenail data:

Parameter QUAD PQL MQL GEE

Intercept group A −1.63 (0.44) −0.72 (0.24) −0.56 (0.17) −0.72 (0.17)

Intercept group B −1.75 (0.45) −0.72 (0.24) −0.53 (0.17) −0.65 (0.17)

Slope group A −0.40 (0.05) −0.29 (0.03) −0.17 (0.02) −0.14 (0.03)

Slope group B −0.57 (0.06) −0.40 (0.04) −0.26 (0.03) −0.25 (0.04)

Var. random intercepts (τ 2) 15.99 (3.02) 4.71 (0.60) 2.49 (0.29)

• Conclusion:

|GEE| < |MQL| < |PQL| < |QUAD|

Part V

Non-linear Models

Chapter 14

Non-Linear Mixed Models

. From linear to non-linear models

. Orange Tree Example

. Song Bird Example

14.1 From Linear to Non-linear Models

• We have studied:

. linear models ←→ generalized linear models

. for Gaussian data ←→ non-Gaussian data

. for cross-sectional (univariate) data ←→ for correlated data (clustered data,multivariate data, longitudinal data)

• In all cases, a certain amount of linearity is preserved:

. E.g., in generalized linear models, linearity operates at the level of the linearpredictor, describing a transformation of the mean (logit, log, probit,. . . )

• This implies that all predictor functions are linear in the parameters:

β0 + β1x1i + . . . βpxpi

• This may still be considered a limiting factor, and non-linearity in the parametricfunctions may be desirable.

• Examples:

. Certain growth curves, especially when they include both a growth spurt aswell as asymptote behavior towards the end of growth

. Dose-response modeling

. Pharmacokinetic and pharmacodynamic models

14.2 LMM and GLMM

• In linear mixed models, the mean is modeled as a linear function of regressionparameters and random effects:

E(Yij|bi) = xij′β + zij

• In generalized linear mixed models, apart from a link function, the mean is againmodeled as a linear function of regression parameters and random effects:

E(Yij|bi) = h(xij′β + zij

′bi)

• In some applications, models are needed, in which the mean is no longer modeledas a function of a linear predictor xij

′β + zij′bi. These are called non-linear

mixed models.

14.3 Non-linear Mixed Models

• In non-linear mixed models, it is assumed that the conditional distribution of Yij,given bi is belongs to the exponential family (Normal, Binomial, Poisson,. . . ).

• The mean is modeled as:

E(Yij|bi) = h(xij,β, zij, bi)

• As before, the random effects are assumed to be normally distributed, with mean0 and covariance D.

• Let fij(yij|bi,β, φ) be the conditional density of Yij given bi, and let f (bi|D) bethe density of the N (0, D) distribution.

• Under conditional independence, the marginal likelihood equals

L(β,D, φ) =N∏

i=1fi(yi|β,D, φ)

∫ ni∏

• The above likelihood is of the same form as the likelihood obtained earlier for ageneralized linear mixed model.

• As before, numerical quadrature is used to approximate the integral

• Non-linear mixed models can also be fitted within the SAS procedure NLMIXED.

14.4 Example: Orange Trees

• We consider an experiment in which the trunk circumference (in mm) is measuredfor 5 orange trees, on 7 different occasions.

• Data:

Response

Day Tree 1 Tree 2 Tree 3 Tree 4 Tree 5

118 30 33 30 32 30

484 58 69 51 62 49

664 87 111 75 112 81

1004 115 156 108 167 125

1231 120 172 115 179 142

1372 142 203 139 209 174

1582 145 203 140 214 177

• The following non-linear mixed model has been proposed:

Yij =β1 + bi

1 + exp[−(tij − β2)/β3]+ εij

bi ∼ N (0, σ2b ) εij ∼ N (0, σ2)

• In SAS PROC NLMIXED, the model can be fitted as follows:

proc nlmixed data=tree;

parms beta1=190 beta2=700 beta3=350

sigmab=10 sigma=10;

num = b;

ex = exp(-(day-beta2)/beta3);

den = 1 + ex;

model y ~ normal(num/den,sigma**2);

random b ~ normal(beta1,sigmab**2) subject=tree;

• An equivalent program is given by

sigmab=10 sigma=10;

num = b + beta1;

den = 1 + ex;

model y ~ normal(num/den,sigma**2);

random b ~ normal(0,sigmab**2) subject=tree;

Specifications

Data Set WORK.TREE

Dependent Variable y

Distribution for Dependent Variable Normal

Random Effects b

Distribution for Random Effects Normal

Subject Variable tree

Optimization Technique Dual Quasi-Newton

Integration Method Adaptive Gaussian

Quadrature

Parameter Estimates

Standard

Parameter Estimate Error DF t Value Pr > |t|

beta1 192.05 15.6577 4 12.27 0.0003

beta2 727.91 35.2487 4 20.65 <.0001

beta3 348.07 27.0798 4 12.85 0.0002

sigmab 31.6463 10.2614 4 3.08 0.0368

sigma 7.8430 1.0125 4 7.75 0.0015

• Note that the number of quadrature points was selected adaptively to be equal toonly one. Refitting the model with 50 quadrature points yielded identical results.

• Empirical Bayes estimates, and subject-specific predictions can be obtained asfollows:

sigmab=10 sigma=10;

num = b + beta1;

den = 1 + ex;

ratio=num/den;

model y ~ normal(ratio,sigma**2);

random b ~ normal(0,sigmab**2) subject=tree out=eb;

predict ratio out=ratio;

• We can now compare the observed data to the subject-specific predictions

yij =β1 + bi

1 + exp[−(tij − β2)/β3]

14.5 Example: Song Birds

• At the University of Antwerp, a novel in-vivo MRI approach to discern thefunctional characteristics of specific neuronal populations in a strongly connectedbrain circuitry has been established.

• Of particular interest: the song control system (SCS) in songbird brain.

• The high vocal center (HVC), one of the major nuclei in this circuit, containsinterneurons and two distinct types of neurons projecting respectively to thenucleus robustus archistriatalis, RA or to area X.

• Schematically,

14.5.1 The MRI Data

• T1-weighted multi slice SE images were acquired (Van Meir et al 2003).

• After acquisition of a set of control images, MnCl2 was injected in the cannulatedHVC.

• Two sets of 5 coronal slices (one through HVC and RA, one through area X) wereacquired every 15 min for up to 6–7 hours after injection.

• This results in 30 data sets of 10 slices of each bird (5 controls and 5 testosteronetreated).

• The change of relative SI of each time point is calculated from the mean signalintensity determined within the defined regions of interest (area X and RA) and inan adjacent control area.

14.5.2 Initial Study

• The effect of testosterone (T) on the dynamics of Mn2+ accumulation in RA andarea X of female starling has been established.

• This has been done with dynamic ME-MRI in individual birds injected withManganese in their HVC.

• This was done in a 2-stage approach: The individual SI data, determined as themean intensity of a region of interest, were submitted to a sigmoid curve fittingproviding curve parameters which revealed upon a two-way repeated ANOVAanalysis that neurons projecting to RA and those to area X were affecteddifferently by testosterone treatment.

• However, this approach could be less reliable if the fit-parameters were mutuallydependent.

• Thus: an integrated non-linear modeling approach is necessary: the MRI signalintensities (SI) need to be fitted to a non-linear mixed model assuming that the SIfollow a pre-specified distribution depending on a covariate indicating the timeafter MnCl2 injection and parameterized through fixed and random (bird-specific)effects.

• An initial model needs to be proposed and then subsequently simplified.

14.5.3 A Model for RA in the Second Period

• Let RAij be the measurement at time j for bird i.

• The following initial non-linear model is assumed:

RAij =(φ0 + φ1Gi + fi)T

η0+η1Gi+niij

(τ0 + τ1Gi + ti)η0+η1Gi+ni + T η0+η1Gi+ni

+γ0 + γ1Gi + εij

• The following conventions are used:

. Gi is an indicator for group membership (0 for control, 1 for active)

. Tij is a covariate indicating the measurement time

. The fixed effects parameters:

∗ the “intercept” parameters φ0, η0, τ0, and γ0

∗ the group effect parameters φ1, η1, τ1, and γ1

∗ If the latter are simultaneously zero, there is no difference between bothgroups. If at least one of them is (significantly) different from zero, then themodel indicates a difference between both groups.

. There are three bird-specific or random effects, fi, ni, and ti, following atrivariate zero-mean normal and general covariance matrix D

. The residual error terms εij are assumed to be mutually independent andindependent from the random effects, and to follow a N (0, σ2) distribution.

• Thus, the general form of the model has 8 fixed-effects parameters, and 7 variancecomponents (3 variances in D, 3 covariances in D, and σ2).

• SAS program:

data hulp2;

set m.vincent03;

if time <= 0 then delete;

if periode = 1 then delete;

proc nlmixed data=hulp2 qpoints=3;

parms phim=0.64 phimdiff=0 eta=1.88 etadiff=0 tau=3.68

taudiff=0 gamma=-0.01 gdiff=0

d11=0.01 sigma2=0.01 d22=0.01 d33=0.01;

teller = (phim + phimdiff * groep + vm ) *

(time ** (eta + etadiff * groep + n ));

noemer = ((tau + t + taudiff * groep )

** (eta + etadiff * groep +n) )

+ (time ** (eta + etadiff * groep + n));

gemid = teller/noemer + gamma + gdiff * groep;

model si_ra ~ normal(gemid,sigma2);

random vm t n ~ normal([0, 0, 0],[d11,0,d22, 0, 0, d33])

subject=vogel out=m.eb;

• Gaussian quadrature is used.

• The covariances in D are set equal to zero, due to computational difficulty.

• Fitting the model produces the following output.

• First, the bookkeeping information is provided:

Specifications

Data Set WORK.HULP2

Dependent Variable SI_RA

Distribution for Dependent Variable Normal

Random Effects vm t n

Distribution for Random Effects Normal

Subject Variable vogel

Optimization Technique Dual Quasi-Newton

Integration Method Adaptive Gaussian

Quadrature

Dimensions

Observations Used 262

Observations Not Used 58

Total Observations 320

Subjects 10

Max Obs Per Subject 28

Parameters 12

Quadrature Points 3

• Next, starting values and the corresponding likelihood are provided:

Parameters

phim phimdiff eta etadiff tau taudiff gamma gdiff d11

0.64 0 1.88 0 3.68 0 -0.01 0 0.01

sigma2 d22 d33 NegLogLike

0.01 0.01 0.01 -325.62539

• The iteration history tells convergence is not straightforward:

Iteration History

Iter Calls NegLogLike Diff MaxGrad Slope

1 14 -571.60061 245.9752 346172 -1341908

2 18 -590.08646 18.48585 25374.59 -3290.57

3 19 -617.20333 27.11687 23321.39 -61.3569

4 21 -630.54411 13.34079 248131.3 -15.4346

5 22 -640.39285 9.848737 7211.013 -21.8031

6 23 -656.41057 16.01772 12295.64 -10.1471

7 25 -666.68301 10.27244 96575.61 -11.8592

8 26 -670.60017 3.917164 100677.4 -5.38138

9 27 -676.45628 5.856102 30099.22 -6.11815

10 30 -677.36448 0.908204 123312.7 -1.98302

20 53 -688.60175 4.468022 32724.37 -4.48668

30 74 -695.15573 0.716814 43575.75 -0.18929

40 93 -697.90828 0.303986 1094.634 -0.51907

50 112 -699.51124 0.002001 166.3506 -0.00338

60 137 -700.79832 0.002657 710.6714 -0.00044

70 157 -701.17648 0.000123 20.20774 -0.00018

71 159 -701.17688 0.000396 125.3658 -0.00006

72 162 -701.21656 0.039676 1426.738 -0.0008

73 164 -701.31217 0.095616 136.553 -0.05932

74 166 -701.31463 0.002454 98.78744 -0.00443

75 168 -701.3147 0.000071 3.915711 -0.00015

76 170 -701.3147 9.862E-7 1.290999 -2.05E-6

NOTE: GCONV convergence criterion satisfied.

• The fit statistics can be used for model comparison, as always:

Fit Statistics

-2 Log Likelihood -1403

AIC (smaller is better) -1379

AICC (smaller is better) -1377

BIC (smaller is better) -1375

• Finally, parameter estimates are provided:

Parameter Estimates

Standard

Parameter Estimate Error DF t Value Pr > |t| Alpha Lower Upper Gradient

phim 0.4085 0.06554 7 6.23 0.0004 0.05 0.2535 0.5634 0.000428

phimdiff 0.08167 0.09233 7 0.88 0.4058 0.05 -0.1366 0.3000 0.000591

eta 2.2117 0.1394 7 15.87 <.0001 0.05 1.8821 2.5412 -9.87E-6

etadiff 0.4390 0.1865 7 2.35 0.0508 0.05 -0.00206 0.8801 0.000717

tau 2.8006 0.2346 7 11.94 <.0001 0.05 2.2457 3.3554 -0.00028

taudiff 0.07546 0.3250 7 0.23 0.8231 0.05 -0.6932 0.8441 0.000046

gamma 0.000237 0.004391 7 0.05 0.9585 0.05 -0.01015 0.01062 0.008095

gdiff 0.001919 0.005923 7 0.32 0.7554 0.05 -0.01209 0.01592 -0.00244

d11 0.02059 0.009283 7 2.22 0.0620 0.05 -0.00136 0.04254 -0.00547

sigma2 0.000180 0.000017 7 10.66 <.0001 0.05 0.000140 0.000221 -1.291

d22 0.2400 0.1169 7 2.05 0.0791 0.05 -0.03633 0.5163 -0.00053

d33 0.02420 0.03488 7 0.69 0.5101 0.05 -0.05829 0.1067 -0.00059

• Next, a backward selection is conducted, using likelihood ratio tests.

• First, the variance d33 of ni was removed. The corresponding test statistic has aχ2

0:1 null distribution (p = 0.4387).

• Next, fixed-effect parameters γ0, γ1, τ1, and φ1 are removed.

• The final model is:

RAij =(φ0 + fi)T

η0+η1Giij

(τ0 + ti)η0+η1Gi + T η0+η1Gi

+ εij

• Overview of parameter estimates and standard errors for initial and final model:

Estimate (s.e.)

Effect Parameter Initial Final

φ0 0.4085(0.0655) 0.4526(0.0478)

φ1 0.0817(0.0923)

η0 2.2117(0.1394) 2.1826(0.0802)

η1 0.4390(0.1865) 0.4285(0.1060)

τ0 2.8006(0.2346) 2.8480(0.1761)

τ1 0.0755(0.3250)

γ0 0.00024(0.0044)

γ1 0.0019(0.0059)

Var(fi) d11 0.0206(0.0092) 0.0225(0.0101)

Var(ti) d22 0.2400(0.1169) 0.2881(0.1338)

Var(ni) d33 0.0242(0.0349)

Var(εij) σ2 0.00018(0.000017) 0.00019(0.000017)

14.5.4 AREA.X at the Second Period

• The same initial model will be fitted to AREAij.

• In this case, a fully general model can be fitted, including also the covariancesbetween the random effects.

• Code to do so:

proc nlmixed data=hulp2 qpoints=3;

parms phim=0.1124 phimdiff=0.1200 eta=2.4158 etadiff=-0.1582 tau=3.8297

taudiff=-0.05259 gamma=-0.00187 gdiff=0.002502 sigma2=0.000156

d11=0.002667 d22=0.2793 d33=0.08505 d12=0 d13=0 d23=0;

teller = (phim + phimdiff * groep + vm)

* (time ** (eta + etadiff * groep + n ));

noemer = ((tau + taudiff * groep +t )

** (eta + etadiff * groep + n) )

+ (time ** (eta + etadiff * groep +n));

gemid = teller/noemer + gamma + gdiff * groep;

model si_area_X ~ normal(gemid,sigma2);

random vm t n ~ normal([0, 0, 0],[d11, d12, d22, d13, d23, d33])

subject=vogel out=m.eb2;

• Model simplification:

. First, the random ni effect is removed (implying the removal of d11, d12, andd13), using a likelihood ratio test statistic with value 4.08 and null distributionχ2

2:3. The corresponding p = 0.1914.

. Removal of the random ti effect is not possible since the likelihood ratio equals54.95 on 1:2 degrees of freedom (p < 0.0001).

. Removal of the covariance between the random ti and fi effects is not possible(G2 = 4.35 on 1 d.f., p = 0.0371).

. Next, the following fixed-effect parameters were removed: γ0, γ1, η1 and τ1.

. The fixed-effect φ1 was found to be highly significant and therefore could notbe removed from the model (G2 = 10.5773 on 1 d.f., p = 0.0011). Thisindicates a significant difference between the two groups.

• The resulting final model is:

AREAij =(φ0 + φ1Gi + fi)T

(τ0 + ti)η0 + T η0

+ εij.

Estimate (s.e.)

Effect Parameter Initial Final

φ0 0.1118 (0.0333) 0.1035 (0.0261)

φ1 0.1116 (0.0458) 0.1331 (0.0312)

η0 2.4940 (0.5390) 2.3462 (0.1498)

η1 -0.0623 (0.5631)

τ0 3.6614 (0.5662) 3.7264 (0.3262)

τ1 -0.1303 (0.6226)

γ0 -0.0021 (0.0032)

γ1 0.0029 (0.0048)

Var(fi) d11 0.0038 (0.0020) 0.004271 (0.0022)

Var(ti) d22 0.2953 (0.2365) 0.5054 (0.2881)

Var(ni) d33 0.1315 (0.1858)

Cov(fi, ti) d12 0.0284 (0.0205) 0.03442 (0.0229)

Cov(fi, ni) d13 -0.0116 (0.0159)

Cov(ti, ni) d23 -0.0095 (0.1615)

Var(εij) σ2 0.00016 (0.000014) 0.00016 (0.000014)

Individual curves, showingdata points as well as empir-ical Bayes predictions for thebird-specific curves, showthat the sigmoidal curvesdescribe the data quite well.

0 1 2 3 4 5 6 7

0 1 2 3 4 5 6 70

0 2 4 6

Group 0Group 1Marg. Group 0Marg. Goup 1

• We can also explore all individualas well as marginal average fittedcurves per group. The marginal effectwas obtained using the sampling-basedmethod.

• It is clear that Area.X is higher for mosttreated birds (group 1) compared tothe untreated birds (group 0).

Part VI

Incomplete Data

Chapter 15

Setting The Scene

. Orthodontic growth data

. Depression trial

. Age-related macular degeneration trial

. Analgesic trial

. Notation

. Taxonomies

15.1 Growth Data

• Taken from Potthoff and Roy, Biometrika (1964)

• Research question:

Is dental growth related to gender ?

• The distance from the center of the pituitary to the maxillary fissure was recordedat ages 8, 10, 12, and 14, for 11 girls and 16 boys

. Much variability between girls / boys

. Considerable variability within girls / boys

. Fixed number of measurements per subject

. Measurements taken at fixed time points

15.2 The Depression Trial

• Clinical trial: experimental drug versus standard drug

• 170 patients

• Response: change versus baseline in HAMD17 score

• 5 post-baseline visits: 4–8

=Visit

4 5 6 7 8

Standard DrugExperimental Drug

15.3 Age-related Macular Degeneration Trial

• Pharmacological Therapy for Macular Degeneration Study Group (1997)

• An occular pressure disease which makes patients progressively lose vision

• 240 patients enrolled in a multi-center trial (190 completers)

• Treatment: Interferon-α (6 million units) versus placebo

• Visits: baseline and follow-up at 4, 12, 24, and 52 weeks

• Continuous outcome: visual acuity: # letters correctly read on a vision chart

• Binary outcome: visual acuity versus baseline ≥ 0 or ≤ 0

• Missingness:

4 wks 12 wks 24 wks 52 wks Number %

Completers

O O O O 188 78.33

Dropouts

O O O M 24 10.00

O O M M 8 3.33

O M M M 6 2.50

M M M M 6 2.50

O O M O 4 1.67

O M M O 1 0.42

M O O O 2 0.83

M O M M 1 0.42

15.4 The Analgesic Trial

• single-arm trial with 530 patients recruited (491 selected for analysis)

• analgesic treatment for pain caused by chronic nonmalignant disease

• treatment was to be administered for 12 months

• we will focus on Global Satisfaction Assessment (GSA)

• GSA scale goes from 1=very good to 5=very bad

• GSA was rated by each subject 4 times during the trial, at months 3, 6, 9, and 12.

• Research questions:

. Evolution over time

. Relation with baseline covariates: age, sex, duration of the pain, type of pain,disease progression, Pain Control Assessment (PCA), . . .

. Investigation of dropout

• Frequencies:

GSA Month 3 Month 6 Month 9 Month 12

1 55 14.3% 38 12.6% 40 17.6% 30 13.5%

2 112 29.1% 84 27.8% 67 29.5% 66 29.6%

3 151 39.2% 115 38.1% 76 33.5% 97 43.5%

4 52 13.5% 51 16.9% 33 14.5% 27 12.1%

5 15 3.9% 14 4.6% 11 4.9% 3 1.4%

Tot 385 302 227 223

• Missingness:

Month 3 Month 6 Month 9 Month 12 Number %

Completers

O O O O 163 41.2

Dropouts

O O O M 51 12.91

O O M M 51 12.91

O M M M 63 15.95

O O M O 30 7.59

O M O O 7 1.77

O M O M 2 0.51

O M M O 18 4.56

M O O O 2 0.51

M O O M 1 0.25

M O M O 1 0.25

M O M M 3 0.76

15.5 LOCF, CC, or Direct Likelihood?

Data:20 30

LOCF:20 30

0 25=⇒

10 65=⇒ θ = 95

200= 0.475 [0.406; 0.544] (biased & too narrow)

CC:20 30

0 0=⇒

10 40=⇒ θ = 20

100= 0.200 [0.122; 0.278] (biased & too wide)

d.l.(MAR):20 30

5 20=⇒

15 60=⇒ θ = 50

200= 0.250 [0.163; 0.337]

15.6 Notation

• Subject i at occasion (time) j = 1, . . . , ni

•Measurement Yij

•Missingness indicator Rij =

1 if Yij is observed,

0 otherwise.

• Group Yij into a vector Y i = (Yi1, . . . , Yini)′ = (Y o

i ,Ymi )

Y oi contains Yij for which Rij = 1,

Y mi contains Yij for which Rij = 0.

• Group Rij into a vector Ri = (Ri1, . . . , Rini)′

• Di: time of dropout: Di = 1 + ∑nij=1Rij

15.7 Modeling Frameworks & Missing Data Mechanisms

f (yi, ri|Xi,θ,ψ)

Selection Models: f (yi|Xi, θ) f (ri|Xi,yoi ,y

mi ,ψ)

MCAR −→ MAR −→ MNAR

f (ri|Xi,ψ) f (ri|Xi,yoi ,ψ) f (ri|Xi,y

mi ,ψ)

Pattern-mixture Models: f (yi|Xi,ri,θ) f (ri|Xi,ψ)

Shared-parameter Models: f (yi|Xi, bi, θ) f (ri|Xi, bi,ψ)

15.8 Frameworks and Their Methods

f (Y i, ri|Xi,θ,ψ)

Selection Models: f (Y i|Xi, θ) f (ri|Xi,Yoi ,Y

mi ,ψ)

MCAR/simple −→ MAR −→ MNAR

CC? direct likelihood! joint model!?

LOCF? direct Bayesian! sensitivity analysis?!

single imputation? multiple imputation (MI)!... IPW ⊃ W-GEE!

d.l. + IPW = double robustness! (consensus)

15.9 Frameworks and Their Methods: Start

mi ,ψ)

direct likelihood!

direct Bayesian!

multiple imputation (MI)!

IPW ⊃ W-GEE!

d.l. + IPW = double robustness!

15.10 Frameworks and Their Methods: Next

mi ,ψ)

joint model!?

sensitivity analysis!

MI (MGK, J&J)

local influence

interval ignorance

IPW based

15.11 Overview

MCAR/simple CC biased

LOCF inefficient

single imputation not simpler than MAR methods

MAR direct lik./Bayes easy to conduct

IPW/d.r. Gaussian & non-Gaussian

multiple imputation

MNAR variety of methods strong, untestable assumptions

most useful in sensitivity analysis

15.12 Ignorability: Direct Likelihood Maximization

MAR : f (Y oi |Xi, θ) f (Di|Xi,Y

oi ,ψ)

Mechanism is MAR

θ and ψ distinct

Interest in θ

Use observed information matrix

=⇒ Likelihood inference is valid

Outcome type Modeling strategy Software

Gaussian Linear mixed model SAS proc MIXED

Non-Gaussian Generalized linear mixed model SAS proc GLIMMIX, NLMIXED

15.12.1 Original, Complete Orthodontic Growth Data

Mean Covar # par

1 unstructured unstructured 18

2 6= slopes unstructured 14

3 = slopes unstructured 13

7 6= slopes CS 6

15.12.2 Trimmed Growth Data: Direct Likelihood

Mean Covar # par

7 6= slopes CS 6

Principle Method Boys at Age 8 Boys at Age 10

Original Direct likelihood, ML 22.88 (0.56) 23.81 (0.49)

Direct likelihood, REML ≡ MANOVA 22.88 (0.58) 23.81 (0.51)

ANOVA per time point 22.88 (0.61) 23.81 (0.53)

Direct Lik. Direct likelihood, ML 22.88 (0.56) 23.17 (0.68)

Direct likelihood, REML 22.88 (0.58) 23.17 (0.71)

MANOVA 24.00 (0.48) 24.14 (0.66)

CC Direct likelihood, ML 24.00 (0.45) 24.14 (0.62)

LOCF Direct likelihood, ML 22.88 (0.56) 22.97 (0.65)

15.12.3 Behind the Scenes

• R completers ↔ N −R “incompleters”

σ11 σ12

• Conditional densityYi2|yi1 ∼ N (β0 + β1yi1, σ22.1)

µ1 freq. & lik. µ1 =1

i=1yi1

µ2 frequentist µ2 =1

i=1yi2

µ2 likelihood µ2 =1

i=1yi2 +

[y2 + β1(yi1 − y1)

15.12.4 Growth Data: Further Comparison of Analyses

Principle Method Boys at Age 8 Boys at Age 10

Original Direct likelihood, ML 22.88 (0.56) 23.81 (0.49)

Direct Lik. Direct likelihood, ML 22.88 (0.56) 23.17 (0.68)

CC Direct likelihood, ML 24.00 (0.45) 24.14 (0.62)

LOCF Direct likelihood, ML 22.88 (0.56) 22.97 (0.65)

Data Mean Covariance Boys at Age 8 Boys at Age 10

Complete Unstructured Unstructured 22.88 23.81

Unstructured CS 22.88 23.81

Unstructured Independence 22.88 23.81

Incomplete Unstructured Unstructured 22.88 23.17

Unstructured CS 22.88 23.52

Unstructured Independence 22.88 24.14

15.12.5 Growth Data: SAS Code for Model 1

IDNR AGE SEX MEASURE

1 8 2 21.0

1 10 2 20.0

1 12 2 21.5

1 14 2 23.0

3 8 2 20.5

3 12 2 24.5

3 14 2 26.0

• SAS code:

proc mixed data = growth method = ml;

class sex idnr age;

model measure = sex age*sex / s;

repeated age / type = un

subject = idnr;

. Subjects in terms of IDNR blocks

. age ensures proper ordering of observationswithin subjects!

15.12.6 Growth Data: SAS Code for Model 2

IDNR AGE SEX MEASURE

1 8 2 21.0

1 10 2 20.0

1 12 2 21.5

1 14 2 23.0

3 8 2 20.5

3 12 2 24.5

3 14 2 26.0

• SAS code:

data help;

set growth;

agecat = age;

proc mixed data = growth method = ml;

class sex idnr agecat;

model measure = sex age*sex / s;

repeated agecat / type = un

subject = idnr;

. Time ordering variable needs to be cate-gorical

15.13 Analysis of the Depression Trial

• Complete case analysis:

%cc(data=depression, id=patient, time=visit, response=change, out={cc});

⇒ performs analysis on CC data set

• LOCF analysis:

%locf(data=depression, id=patient, time=visit, response=change, out={locf});

⇒ performs analysis on LOCF data

• Direct-likelihood analysis: ⇒ fit linear mixed model to incomplete data

• Treatment effect at visit 8 (last follow-up measurement):

Method Estimate (s.e.) p-value

CC -1.94 (1.17) 0.0995

LOCF -1.63 (1.08) 0.1322

MAR -2.38 (1.16) 0.0419

Observe the slightly significant p-value under the MAR model

15.14 Weighted Generalized Estimating Equations

. General Principle

. Analysis of the analgesic trial

. Analysis of the ARMD trial

. Analysis of the depression trial

15.15 General Principle

MAR and non-ignorable !

• Standard GEE inference correct only under MCAR

Under MAR: weighted GEERobins, Rotnitzky & Zhao (JASA, 1995)

Fitzmaurice, Molenberghs & Lipsitz (JRSSB, 1995)

• Decompose dropout time Di = (Ri1, . . . , Rin) = (1, . . . , 1, 0, . . . , 0)

• Weigh a contribution by inverse dropout probability

νidi≡ P [Di = di] =

di−1∏

k=2(1− P [Rik = 0|Ri2 = . . . = Ri,k−1 = 1]) ×

P [Ridi= 0|Ri2 = . . . = Ri,di−1 = 1]I{di≤T}

• Adjust estimating equations

· ∂µi∂β′

V −1i (yi − µi) = 0

15.16 Computing the Weights

• Predicted values from (PROC GENMOD) output

• The weights are now defined at the individual measurement level:

. At the first occasion, the weight is w = 1

. At other than the last ocassion, the weight is the already accumulated weight,multiplied by 1−the predicted probability

. At the last occasion within a sequence where dropout occurs the weight ismultiplied by the predicted probability

. At the end of the process, the weight is inverted

15.17 Analysis of the Analgesic Trial

• A logistic regression for the dropout indicator:

logit[P (Di = j|Di ≥ j, ·)] = ψ0 + ψ11I(GSAi,j−1 = 1) + ψ12I(GSAi,j−1 = 2)

+ψ13I(GSAi,j−1 = 3) + ψ14I(GSAi,j−1 = 4)

+ψ2PCA0i + ψ3PFi + ψ4GDi

. GSAi,j−1 the 5-point outcome at the previous time

. I(·) is an indicator function

. PCA0i is pain control assessment at baseline

. PFi is physical functioning at baseline

. GDi is genetic disorder at baseline are used

Effect Par. Estimate (s.e.)

Intercept ψ0 -1.80 (0.49)

Previous GSA= 1 ψ11 -1.02 (0.41)

Previous GSA= 2 ψ12 -1.04 (0.38)

Previous GSA= 3 ψ13 -1.34 (0.37)

Previous GSA= 4 ψ14 -0.26 (0.38)

Basel. PCA ψ2 0.25 (0.10)

Phys. func. ψ3 0.009 (0.004)

Genetic disfunc. ψ4 0.59 (0.24)

• There is some evidence for MAR: P (Di = j|Di ≥ j) depends on previous GSA.

• Furthermore: baseline PCA, physical functioning and genetic/congenital disorder.

• GEE and WGEE:

logit[P (Yij = 1|tj,PCA0i)] = β1 + β2tj + β3t2j + β4PCA0i

Effect Par. GEE WGEE

Intercept β1 2.95 (0.47) 2.17 (0.69)

Time β2 -0.84 (0.33) -0.44 (0.44)

Time2 β3 0.18 (0.07) 0.12 (0.09)

Basel. PCA β4 -0.24 (0.10) -0.16 (0.13)

• A hint of potentially important differences between both

15.18 Analgesic Trial: Steps for WGEE in SAS

1. Preparatory data manipulation:

%dropout(...)

2. Logistic regression for weight model:

proc genmod data=gsac;

class prevgsa;

model dropout = prevgsa pca0 physfct gendis / pred dist=b;

ods output obstats=pred;

3. Conversion of predicted values to weights:

%dropwgt(...)

4. Weighted GEE analysis:

proc genmod data=repbin.gsaw;

scwgt wi;

class patid timecls;

model gsabin = time|time pca0 / dist=b;

repeated subject=patid / type=un corrw within=timecls;

15.19 Analysis of the ARMD Trial

• Model for the weights:

logit[P (Di = j|Di ≥ j)] = ψ0 + ψ1yi,j−1 + ψ2Ti + ψ31L1i + ψ32L2i + ψ34L3i

+ψ41I(tj = 2) + ψ42I(tj = 3)

. yi,j−1 the binary outcome at the previous time ti,j−1 = tj−1 (since time iscommon to all subjects)

. Ti = 1 for interferon-α and Ti = 0 for placebo

. Lki = 1 if the patient’s eye lesion is of level k = 1, . . . , 4 (since one dummyvariable is redundant, only three are used)

. I(·) is an indicator function

• Results for the weights model:

Effect Parameter Estimate (s.e.)

Intercept ψ0 0.14 (0.49)

Previous outcome ψ1 0.04 (0.38)

Treatment ψ2 -0.86 (0.37)

Lesion level 1 ψ31 -1.85 (0.49)

Lesion level 2 ψ32 -1.91 (0.52)

Lesion level 3 ψ33 -2.80 (0.72)

Time 2 ψ41 -1.75 (0.49)

Time 3 ψ42 -1.38 (0.44)

• GEE:logit[P (Yij = 1|Ti, tj)] = βj1 + βj2Ti

. Ti = 0 for placebo and Ti = 1 for interferon-α

. tj (j = 1, . . . , 4) refers to the four follow-up measurements

. Classical GEE and linearization-based GEE

. Comparison between CC, LOCF, and GEE analyses

• SAS code: Molenberghs and Verbeke (2005, Section 32.5)

• Results:

Effect Par. CC LOCF Observed data

Unweighted WGEE

Int.4 β11 -1.01(0.24;0.24) -0.87(0.20;0.21) -0.87(0.21;0.21) -0.98(0.10;0.44)

Int.12 β21 -0.89(0.24;0.24) -0.97(0.21;0.21) -1.01(0.21;0.21) -1.78(0.15;0.38)

Int.24 β31 -1.13(0.25;0.25) -1.05(0.21;0.21) -1.07(0.22;0.22) -1.11(0.15;0.33)

Int.52 β41 -1.64(0.29;0.29) -1.51(0.24;0.24) -1.71(0.29;0.29) -1.72(0.25;0.39)

Tr.4 β12 0.40(0.32;0.32) 0.22(0.28;0.28) 0.22(0.28;0.28) 0.80(0.15;0.67)

Tr.12 β22 0.49(0.31;0.31) 0.55(0.28;0.28) 0.61(0.29;0.29) 1.87(0.19;0.61)

Tr.24 β32 0.48(0.33;0.33) 0.42(0.29;0.29) 0.44(0.30;0.30) 0.73(0.20;0.52)

Tr.52 β42 0.40(0.38;0.38) 0.34(0.32;0.32) 0.44(0.37;0.37) 0.74(0.31;0.52)

Corr. ρ 0.39 0.44 0.39 0.33

15.20 Analysis of the Depression Trial

• Response: create binary indicator ybin for HAMD17 > 7

• Model for dropout:

logit[P (Di = j|Di ≥ j)] = ψ0 + ψ1yi,j−1 + γTi

. yi,j−1: the binary indicator at the previous occasion

. Ti: treatment indicator for patient i

• SAS code:

. Preparing the dataset:

%dropout(data=depression,id=patient,time=visit,response=ybin,out=dropout);

producing:

. dropout indicates whether missingness at a given time occurs

. prev contains outcome at the previous occasion

. The logistic model for dropout:

proc genmod data=dropout descending;

class trt;

model dropout = prev trt / pred dist=b;

output out=pred p=pred;

. The weights can now be included in the GENMOD program which specifies theGEE, through the WEIGHT or SCWGT statements:

proc genmod data=study descending;

weight wi;

class patient visitclass trt;

model ybin = trt visit trt*visit basval basval*visit / dist=bin;

repeated subject=patient / withinsubject=visitclass type=cs corrw;

• Results:

WGEE GEE

Effect est. (s.e.) p-value est. (s.e.) p-value

Treatment at visit 4 -1.57 (0.99) 0.11 -0.24 (0.57) 0.67

Treatment at visit 5 -0.67 (0.65) 0.30 0.09 (0.40) 0.82

Treatment at visit 6 0.62 (0.56) 0.27 0.17 (0.34) 0.62

15.21 Multiple Imputation

. General idea

. Estimation

. Hypothesis testing

. Use of MI in practice

. Analysis of the growth data

. Analysis of the ARMD trial

. Creating monotone missingness

15.22 General Principles

• Valid under MAR

• An alternative to direct likelihood and WGEE

• Three steps:

1. The missing values are filled in M times =⇒ M complete data sets

2. The M complete data sets are analyzed by using standard procedures

3. The results from the M analyses are combined into a single inference

• Rubin (1987), Rubin and Schenker (1986), Little and Rubin (1987)

15.23 Use of MI in Practice

• Many analyses of the same incomplete set of data

• A combination of missing outcomes and missing covariates

• As an alternative to WGEE: MI can be combined with classical GEE

• MI in SAS:

Imputation Task: PROC MI

Analysis Task: PROC “MYFAVORITE”

Inference Task: PROC MIANALYZE

15.24 MI Analysis of the ARMD Trial

•M = 10 imputations

• GEE:logit[P (Yij = 1|Ti, tj)] = βj1 + βj2Ti

• GLMM:

logit[P (Yij = 1|Ti, tj, bi)] = βj1 + bi + βj2Ti, bi ∼ N (0, τ 2)

• Ti = 0 for placebo and Ti = 1 for interferon-α

• tj (j = 1, . . . , 4) refers to the four follow-up measurements

• Imputation based on the continuous outcome

• Results:

Effect Par. GEE GLMM

Int.4 β11 -0.84(0.20) -1.46(0.36)

Int.12 β21 -1.02(0.22) -1.75(0.38)

Int.24 β31 -1.07(0.23) -1.83(0.38)

Int.52 β41 -1.61(0.27) -2.69(0.45)

Trt.4 β12 0.21(0.28) 0.32(0.48)

Trt.12 β22 0.60(0.29) 0.99(0.49)

Trt.24 β32 0.43(0.30) 0.67(0.51)

Trt.52 β42 0.37(0.35) 0.52(0.56)

R.I. s.d. τ 2.20(0.26)

R.I. var. τ 2 4.85(1.13)

15.25 SAS Code for MI

1. Preparatory data analysis so that there is one line per subject

2. The imputation task:

proc mi data=armd13 seed=486048 out=armd13a simple nimpute=10 round=0.1;

var lesion diff4 diff12 diff24 diff52;

by treat;

Note that the imputation task is conducted on the continuous outcome ‘diff·’,indicating the difference in number of letters versus baseline

3. Then, data manipulation takes place to define the binary indicators and to createa longitudinal version of the dataset

4. The analysis task (GEE):

proc genmod data=armd13c;

class time subject;

by _imputation_;

model bindif = time1 time2 time3 time4

trttime1 trttime2 trttime3 trttime4

/ noint dist=binomial covb;

repeated subject=subject / withinsubject=time type=exch modelse;

ods output ParameterEstimates=gmparms parminfo=gmpinfo CovB=gmcovb;

5. The analysis task (GLMM):

proc nlmixed data=armd13c qpoints=20 maxiter=100 technique=newrap cov ecov;

by _imputation_;

eta = beta11*time1+beta12*time2+beta13*time3+beta14*time4+b

+beta21*trttime1+beta22*trttime2+beta23*trttime3+beta24*trttime4;

p = exp(eta)/(1+exp(eta));

model bindif ~ binary(p);

random b ~ normal(0,tau*tau) subject=subject;

estimate ’tau2’ tau*tau;

ods output ParameterEstimates=nlparms CovMatParmEst=nlcovb

AdditionalEstimates=nlparmsa CovMatAddEst=nlcovba;

6. The inference task (GEE):

proc mianalyze parms=gmparms covb=gmcovb parminfo=gmpinfo wcov bcov tcov;

modeleffects time1 time2 time3 time4 trttime1 trttime2 trttime3 trttime4;

7. The inference task (GLMM):

proc mianalyze parms=nlparms covb=nlcovb wcov bcov tcov;

modeleffects beta11 beta12 beta13 beta14 beta21 beta22 beta23 beta24;

Chapter 16

Overview

LOCF inefficient

multiple imputation

Part VII

Topics in Methods and Sensitivity Analysis for IncompleteData

Chapter 17

An MNAR Selection Model and Local Influence

. The Diggle and Kenward selection model

. Mastitis in dairy cattle

. An informal sensitivity analysis

. Local influence to conduct sensitivity analysis

17.1 A Full Selection Model

MNAR :∫f (Y i|θ)f (Di|Y i,ψ)dY m

f (Y i|θ)

Linear mixed model

Y i = Xiβ + Zibi + εi

f (Di|Y i,ψ)

Logistic regressions for dropout

logit [P (Di = j | Di ≥ j, Yi,j−1, Yij)]

= ψ0 + ψ1Yi,j−1 + ψ2Yij

Diggle and Kenward (JRSSC 1994)

17.2 Mastitis in Dairy Cattle

• Infectious disease of the udder

• Leads to a reduction in milk yield

• High yielding cows more susceptible?

• But this cannot be measured directly be-cause of the effect of the disease: ev-idence is missing since infected causehave no reported milk yield

•Model for milk yield:

µ + ∆

σ21 ρσ1σ2

ρσ1σ2 σ21

•Model for mastitis:

logit [P (Ri = 1|Yi1, Yi2)] = ψ0 + ψ1Yi1 + ψ2Yi2

= 0.37 + 2.25Yi1 − 2.54Yi2

= 0.37 − 0.29Yi1 − 2.54(Yi2 − Yi1)

• LR test for H0 : ψ2 = 0 : G2 = 5.11

17.3 Criticism −→ Sensitivity Analysis

“. . . , estimating the ‘unestimable’ can be accomplished only by makingmodelling assumptions,. . . . The consequences of model misspeci-fication will (. . . ) be more severe in the non-random case.” (Laird1994)

• Change distributional assumptions (Kenward 1998)

• Local and global influence methods

• Pattern-mixture models

• Several plausible models or ranges of inferences

• Semi-parametric framework (Scharfstein et al 1999)

17.4 Kenward’s Sensitivity Analysis

• Deletion of #4 and #5 ⇒ G2 for ψ2:5.11 −→ 0.08

• Cows #4 and #5 have unusually largeincrements

• Kenward conjectures: #4 and #5 ill dur-ing the first year

• Kenward (SiM 1998)

17.5 Local Influence

• Verbeke, Thijs, Lesaffre, Kenward (Bcs 2001)

• Perturbed MAR dropout model:

logit [P (Di = 1|Yi1, Yi2)]

= ψ0 + ψ1Yi1 + ωiYi2

• Likelihood displacement:

LD(ω) = 2[Lω=0

(θ, ψ

)− Lω=0

(θω,

) ]≥ 0

17.5 Local Influence

• Verbeke, Thijs, Lesaffre, Kenward (Bcs 2001)

• Perturbed MAR dropout model:

logit [P (Di = 1|Yi1, Yi2)]

= ψ0 + ψ1Yi1 + ωiYi2

or ψ0 + ψ1Yi1 + ωi(Yi2 − Yi1)

• Likelihood displacement:

LD(ω) = 2[Lω=0

(θ, ψ

)− Lω=0

(θω,

) ]≥ 0

17.6 Application to Mastitis Data

• Removing #4, #5 and #66

⇒ G2 = 0.005

• hmax: different signs for (#4,#5) and #66

Chapter 18

Interval of Ignorance

. The Slovenian Public Opinion Survey

. MAR and MNAR analyses

. Informal sensitivity analysis

. Interval of ignorance & interval of uncertainty

18.1 The Slovenian Plebiscite

• Rubin, Stern, and Vehovar (1995)

• Slovenian Public Opinion (SPO) Survey

• Four weeks prior to decisive plebiscite

• Three questions:

1. Are you in favor of Slovenian independence ?

2. Are you in favor of Slovenia’s secession from Yugoslavia ?

3. Will you attend the plebiscite ?

• Political decision: ABSENCE≡NO

• Primary Estimand: θ: Proportion in favor of independence

• Slovenian Public Opinion Survey Data:

Independence

Secession Attendance Yes No ∗Yes Yes 1191 8 21

No 8 0 4

∗ 107 3 9

No Yes 158 68 29

No 7 14 3

∗ 18 43 31

∗ Yes 90 2 109

No 1 2 25

∗ 19 8 96

18.2 Slovenian Public Opinion: 1st Analysis

• Pessimistic: All who can say NO will say NO

θ =1439

2074= 0.694

• Optimistic: All who can say YES will say YES

θ =1439 + 159 + 144 + 136

2076= 0.904

• Resulting Interval:

θ ∈ [0.694; 0.904]

• Resulting Interval:

θ ∈ [0.694; 0.904]

• Complete cases: All who answered on 3 questions

θ =1191 + 158

1454= 0.928 ?

• Available cases: All who answered on both questions

θ =1191 + 158 + 90

1549= 0.929 ?

18.3 Slovenian Public Opinion: 2nd Analysis

•Missing at Random:

Non-response is allowed to depend on observed, but not on unobserved outcomes:

. Based on two questions:θ = 0.892

. Based on three questions:θ = 0.883

•Missing Not at Random (NI):

Non-response is allowed to depend on unobserved measurements:

θ = 0.782

18.4 Slovenian Public Opinion Survey

Estimator θ

Pessimistic bound 0.694

Optimistic bound 0.904

Complete cases 0.928 ?

Available cases 0.929 ?

MAR (2 questions) 0.892

MNAR 0.782

18.5 Slovenian Plebiscite: The Truth ?

θ =0.885

Estimator θ

Pessimistic bound 0.694

Optimistic bound 0.904

Complete cases 0.928 ?

Available cases 0.929 ?

MNAR 0.782

18.6 Did “the” MNAR model behave badly ?

Consider a family of MNAR models

• Baker, Rosenberger, and DerSimonian (1992)

• Counts Yr1r2jk

• j, k = 1, 2 indicates YES/NO

• r1, r2 = 0, 1 indicates MISSING/OBSERVED

18.6.1 Model Formulation

E(Y11jk) = mjk,

E(Y10jk) = mjkβjk,

E(Y01jk) = mjkαjk,

E(Y00jk) = mjkαjkβjkγjk,

Interpretation:

• αjk: models non-response on independence question

• βjk: models non-response on attendance question

• γjk: interaction between both non-response indicators (cannot depend on j or k)

18.6.2 Identifiable Models

Model Structure d.f. loglik θ C.I.

BRD1 (α, β) 6 -2495.29 0.892 [0.878;0.906]

BRD2 (α, βj) 7 -2467.43 0.884 [0.869;0.900]

BRD3 (αk, β) 7 -2463.10 0.881 [0.866;0.897]

BRD4 (α, βk) 7 -2467.43 0.765 [0.674;0.856]

BRD5 (αj, β) 7 -2463.10 0.844 [0.806;0.882]

BRD6 (αj, βj) 8 -2431.06 0.819 [0.788;0.849]

BRD7 (αk, βk) 8 -2431.06 0.764 [0.697;0.832]

BRD8 (αj, βk) 8 -2431.06 0.741 [0.657;0.826]

BRD9 (αk, βj) 8 -2431.06 0.867 [0.851;0.884]

18.6.3 An “Interval” of MNAR Estimates

θ =0.885

Estimator θ

[Pessimistic; optimistic] [0.694;0.904]

Complete cases 0.928

Available cases 0.929

MNAR 0.782

MNAR “interval” [0.741;0.892]

18.7 A More Formal Look

Statistical Imprecision Statistical Ignorance

Statistical Uncertainty

��

Statistical Imprecision: Due to finite sampling

• Fundamental concept of mathematical statistics

• Consistency, efficiency, precision, testing,. . .

• Disappears as sample size increases

Statistical Ignorance: Due to incomplete observations

• Received less attention

• Can invalidate conclusions

• Does not disappear with increasing sample size

Kenward, Goetghebeur, and Molenberghs (StatMod 2001)

18.8 Slovenian Public Opinion: 3rd Analysis

Model Structure d.f. loglik θ C.I.

BRD1 (α, β) 6 -2495.29 0.892 [0.878;0.906]

BRD2 (α, βj) 7 -2467.43 0.884 [0.869;0.900]

BRD3 (αk, β) 7 -2463.10 0.881 [0.866;0.897]

BRD4 (α, βk) 7 -2467.43 0.765 [0.674;0.856]

BRD5 (αj, β) 7 -2463.10 0.844 [0.806;0.882]

BRD6 (αj, βj) 8 -2431.06 0.819 [0.788;0.849]

BRD7 (αk, βk) 8 -2431.06 0.764 [0.697;0.832]

BRD8 (αj, βk) 8 -2431.06 0.741 [0.657;0.826]

BRD9 (αk, βj) 8 -2431.06 0.867 [0.851;0.884]

Model 10 (αk, βjk) 9 -2431.06 [0.762;0.893] [0.744;0.907]

Model 11 (αjk, βj) 9 -2431.06 [0.766;0.883] [0.715;0.920]

Model 12 (αjk, βjk) 10 -2431.06 [0.694;0.904]

18.9 Every MNAR Model Has Got a MAR Bodyguard

• Fit an MNAR model to a set of incomplete data

• Change the conditional distribution of the unobserved outcomes, given theobserved ones, to comply with MAR

• The resulting new model will have exactly the same fit as the original MNARmodel

• The missing data mechanism has changed

• This implies that definitively testing for MAR versus MNAR is not possible

18.10 Slovenian Public Opinion: 4rd Analysis

Model Structure d.f. loglik θ C.I. θMAR

BRD1 (α, β) 6 -2495.29 0.892 [0.878;0.906] 0.8920

BRD2 (α, βj) 7 -2467.43 0.884 [0.869;0.900] 0.8915

BRD3 (αk, β) 7 -2463.10 0.881 [0.866;0.897] 0.8915

BRD4 (α, βk) 7 -2467.43 0.765 [0.674;0.856] 0.8915

BRD5 (αj, β) 7 -2463.10 0.844 [0.806;0.882] 0.8915

BRD6 (αj, βj) 8 -2431.06 0.819 [0.788;0.849] 0.8919

BRD7 (αk, βk) 8 -2431.06 0.764 [0.697;0.832] 0.8919

BRD8 (αj, βk) 8 -2431.06 0.741 [0.657;0.826] 0.8919

BRD9 (αk, βj) 8 -2431.06 0.867 [0.851;0.884] 0.8919

Model 10 (αk, βjk) 9 -2431.06 [0.762;0.893] [0.744;0.907] 0.8919

Model 11 (αjk, βj) 9 -2431.06 [0.766;0.883] [0.715;0.920] 0.8919

Model 12 (αjk, βjk) 10 -2431.06 [0.694;0.904] 0.8919

θ =0.885

Estimator θ

[Pessimistic; optimistic] [0.694;0.904]

MNAR 0.782

MNAR “interval” [0.753;0.891]

Model 10 [0.762;0.893]

Model 11 [0.766;0.883]

Model 12 [0.694;0.904]

Chapter 19

Overview

LOCF inefficient

multiple imputation

Longitudinal and Incomplete Data - KU Leuven · Longitudinal and Incomplete Data Geert Verbeke...

Documents

Transcript of Longitudinal and Incomplete Data - KU Leuven · Longitudinal and Incomplete Data Geert Verbeke...

Predicting time-to-event outcomes based on high ... events... · high-dimensional multivariate longitudinal information EFSPI: Brussels, November 7, 2013 Geert Verbeke ... • Conditional

Analyzing Incomplete Discrete Longitudinal Clinical Trial Data · LONGITUDINAL CLINICAL TRIAL DATA 3 subject-speciﬁc) and conditional models. We then argue that marginal and random-eﬀects

Geert Berghuis

CURRICULUM VITAE GEERT JAN SOMSEN EDUCATION AND DEGREESfasos.maastrichtuniversity.nl/profiles/Somsen/CV Geert Somsen.pdf · CURRICULUM VITAE GEERT JAN SOMSEN EDUCATION AND DEGREES

Presentation geert de cock kartuzy_fracking_english_final_long

200X, Van Geert, DSA

Geert Van Grootel

Bom-vl Workshop Geert Wissink

Introduction to Longitudinal Data Analysis › ... › aulas › 2014 › IIWorkshop_Longitudinal_Data › … · Introduction to Longitudinal Data Analysis Geert Molenberghs geert.molenberghs@uhasselt.be

An Analytic Road Map for Incomplete Longitudinal Clinical Trial Data Craig Mallinckrodt Graybill Conference June 12, 2008 Fort Collins, CO.

Geert Maria Jan LERNOUT

Jan Haenraets & Geert Robberechts (INTACH)

GEERT VERBEKE – GEERT MOLENBERGHS Longitudinal and …ferran.torres.name/download/shared/missing/51_.pdf · 2020. 4. 3. · 148 GEERT VERBEKE – GEERT MOLENBERGHS Akey issue in

ESALQ Course on Models for Longitudinal and Incomplete ...

Introduction to Longitudinal Data Analysisdms/Longitudinal/geert_lda.pdfIntroduction to Longitudinal Data Analysis Geert Molenberghs Center for Statistics Universiteit Hasselt, Belgium

Geert Klein Breteler

FBE Geert Hofstede Teaching Learning

GEERT GOIRIS

Characterizing Persistent Disturbing Behavior Using Longitudinal and Multivariate Techniques Jan Serroyen, UHasselt Liesbeth Bruckers, UHasselt Geert Rogiers,

IMGENT - Geert Stox